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_UNLOCKED;
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, int sync)
614 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
617 static inline int cpu_of(struct rq *rq)
627 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
628 * See detach_destroy_domains: synchronize_sched for details.
630 * The domain tree of any CPU may only be accessed from within
631 * preempt-disabled sections.
633 #define for_each_domain(cpu, __sd) \
634 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
636 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
637 #define this_rq() (&__get_cpu_var(runqueues))
638 #define task_rq(p) cpu_rq(task_cpu(p))
639 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
641 static inline void update_rq_clock(struct rq *rq)
643 rq->clock = sched_clock_cpu(cpu_of(rq));
647 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
649 #ifdef CONFIG_SCHED_DEBUG
650 # define const_debug __read_mostly
652 # define const_debug static const
658 * Returns true if the current cpu runqueue is locked.
659 * This interface allows printk to be called with the runqueue lock
660 * held and know whether or not it is OK to wake up the klogd.
662 int runqueue_is_locked(void)
665 struct rq *rq = cpu_rq(cpu);
668 ret = spin_is_locked(&rq->lock);
674 * Debugging: various feature bits
677 #define SCHED_FEAT(name, enabled) \
678 __SCHED_FEAT_##name ,
681 #include "sched_features.h"
686 #define SCHED_FEAT(name, enabled) \
687 (1UL << __SCHED_FEAT_##name) * enabled |
689 const_debug unsigned int sysctl_sched_features =
690 #include "sched_features.h"
695 #ifdef CONFIG_SCHED_DEBUG
696 #define SCHED_FEAT(name, enabled) \
699 static __read_mostly char *sched_feat_names[] = {
700 #include "sched_features.h"
706 static int sched_feat_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 __init 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 inline 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_PERCPU;
1129 #else /* CONFIG_SCHED_HRTICK */
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)
1141 #endif /* CONFIG_SCHED_HRTICK */
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->nvcsw | LONG_MIN; /* sets MSB */
1944 task_rq_unlock(rq, &flags);
1947 * If it changed from the expected state, bail out now.
1949 if (unlikely(!ncsw))
1953 * Was it really running after all now that we
1954 * checked with the proper locks actually held?
1956 * Oops. Go back and try again..
1958 if (unlikely(running)) {
1964 * It's not enough that it's not actively running,
1965 * it must be off the runqueue _entirely_, and not
1968 * So if it wa still runnable (but just not actively
1969 * running right now), it's preempted, and we should
1970 * yield - it could be a while.
1972 if (unlikely(on_rq)) {
1973 schedule_timeout_uninterruptible(1);
1978 * Ahh, all good. It wasn't running, and it wasn't
1979 * runnable, which means that it will never become
1980 * running in the future either. We're all done!
1989 * kick_process - kick a running thread to enter/exit the kernel
1990 * @p: the to-be-kicked thread
1992 * Cause a process which is running on another CPU to enter
1993 * kernel-mode, without any delay. (to get signals handled.)
1995 * NOTE: this function doesnt have to take the runqueue lock,
1996 * because all it wants to ensure is that the remote task enters
1997 * the kernel. If the IPI races and the task has been migrated
1998 * to another CPU then no harm is done and the purpose has been
2001 void kick_process(struct task_struct *p)
2007 if ((cpu != smp_processor_id()) && task_curr(p))
2008 smp_send_reschedule(cpu);
2013 * Return a low guess at the load of a migration-source cpu weighted
2014 * according to the scheduling class and "nice" value.
2016 * We want to under-estimate the load of migration sources, to
2017 * balance conservatively.
2019 static unsigned long source_load(int cpu, int type)
2021 struct rq *rq = cpu_rq(cpu);
2022 unsigned long total = weighted_cpuload(cpu);
2024 if (type == 0 || !sched_feat(LB_BIAS))
2027 return min(rq->cpu_load[type-1], total);
2031 * Return a high guess at the load of a migration-target cpu weighted
2032 * according to the scheduling class and "nice" value.
2034 static unsigned long target_load(int cpu, int type)
2036 struct rq *rq = cpu_rq(cpu);
2037 unsigned long total = weighted_cpuload(cpu);
2039 if (type == 0 || !sched_feat(LB_BIAS))
2042 return max(rq->cpu_load[type-1], total);
2046 * find_idlest_group finds and returns the least busy CPU group within the
2049 static struct sched_group *
2050 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2052 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2053 unsigned long min_load = ULONG_MAX, this_load = 0;
2054 int load_idx = sd->forkexec_idx;
2055 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2058 unsigned long load, avg_load;
2062 /* Skip over this group if it has no CPUs allowed */
2063 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2066 local_group = cpu_isset(this_cpu, group->cpumask);
2068 /* Tally up the load of all CPUs in the group */
2071 for_each_cpu_mask_nr(i, group->cpumask) {
2072 /* Bias balancing toward cpus of our domain */
2074 load = source_load(i, load_idx);
2076 load = target_load(i, load_idx);
2081 /* Adjust by relative CPU power of the group */
2082 avg_load = sg_div_cpu_power(group,
2083 avg_load * SCHED_LOAD_SCALE);
2086 this_load = avg_load;
2088 } else if (avg_load < min_load) {
2089 min_load = avg_load;
2092 } while (group = group->next, group != sd->groups);
2094 if (!idlest || 100*this_load < imbalance*min_load)
2100 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2103 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2106 unsigned long load, min_load = ULONG_MAX;
2110 /* Traverse only the allowed CPUs */
2111 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2113 for_each_cpu_mask_nr(i, *tmp) {
2114 load = weighted_cpuload(i);
2116 if (load < min_load || (load == min_load && i == this_cpu)) {
2126 * sched_balance_self: balance the current task (running on cpu) in domains
2127 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2130 * Balance, ie. select the least loaded group.
2132 * Returns the target CPU number, or the same CPU if no balancing is needed.
2134 * preempt must be disabled.
2136 static int sched_balance_self(int cpu, int flag)
2138 struct task_struct *t = current;
2139 struct sched_domain *tmp, *sd = NULL;
2141 for_each_domain(cpu, tmp) {
2143 * If power savings logic is enabled for a domain, stop there.
2145 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2147 if (tmp->flags & flag)
2155 cpumask_t span, tmpmask;
2156 struct sched_group *group;
2157 int new_cpu, weight;
2159 if (!(sd->flags & flag)) {
2165 group = find_idlest_group(sd, t, cpu);
2171 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2172 if (new_cpu == -1 || new_cpu == cpu) {
2173 /* Now try balancing at a lower domain level of cpu */
2178 /* Now try balancing at a lower domain level of new_cpu */
2181 weight = cpus_weight(span);
2182 for_each_domain(cpu, tmp) {
2183 if (weight <= cpus_weight(tmp->span))
2185 if (tmp->flags & flag)
2188 /* while loop will break here if sd == NULL */
2194 #endif /* CONFIG_SMP */
2197 * try_to_wake_up - wake up a thread
2198 * @p: the to-be-woken-up thread
2199 * @state: the mask of task states that can be woken
2200 * @sync: do a synchronous wakeup?
2202 * Put it on the run-queue if it's not already there. The "current"
2203 * thread is always on the run-queue (except when the actual
2204 * re-schedule is in progress), and as such you're allowed to do
2205 * the simpler "current->state = TASK_RUNNING" to mark yourself
2206 * runnable without the overhead of this.
2208 * returns failure only if the task is already active.
2210 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2212 int cpu, orig_cpu, this_cpu, success = 0;
2213 unsigned long flags;
2217 if (!sched_feat(SYNC_WAKEUPS))
2221 if (sched_feat(LB_WAKEUP_UPDATE)) {
2222 struct sched_domain *sd;
2224 this_cpu = raw_smp_processor_id();
2227 for_each_domain(this_cpu, sd) {
2228 if (cpu_isset(cpu, sd->span)) {
2237 rq = task_rq_lock(p, &flags);
2238 old_state = p->state;
2239 if (!(old_state & state))
2247 this_cpu = smp_processor_id();
2250 if (unlikely(task_running(rq, p)))
2253 cpu = p->sched_class->select_task_rq(p, sync);
2254 if (cpu != orig_cpu) {
2255 set_task_cpu(p, cpu);
2256 task_rq_unlock(rq, &flags);
2257 /* might preempt at this point */
2258 rq = task_rq_lock(p, &flags);
2259 old_state = p->state;
2260 if (!(old_state & state))
2265 this_cpu = smp_processor_id();
2269 #ifdef CONFIG_SCHEDSTATS
2270 schedstat_inc(rq, ttwu_count);
2271 if (cpu == this_cpu)
2272 schedstat_inc(rq, ttwu_local);
2274 struct sched_domain *sd;
2275 for_each_domain(this_cpu, sd) {
2276 if (cpu_isset(cpu, sd->span)) {
2277 schedstat_inc(sd, ttwu_wake_remote);
2282 #endif /* CONFIG_SCHEDSTATS */
2285 #endif /* CONFIG_SMP */
2286 schedstat_inc(p, se.nr_wakeups);
2288 schedstat_inc(p, se.nr_wakeups_sync);
2289 if (orig_cpu != cpu)
2290 schedstat_inc(p, se.nr_wakeups_migrate);
2291 if (cpu == this_cpu)
2292 schedstat_inc(p, se.nr_wakeups_local);
2294 schedstat_inc(p, se.nr_wakeups_remote);
2295 update_rq_clock(rq);
2296 activate_task(rq, p, 1);
2300 trace_mark(kernel_sched_wakeup,
2301 "pid %d state %ld ## rq %p task %p rq->curr %p",
2302 p->pid, p->state, rq, p, rq->curr);
2303 check_preempt_curr(rq, p, sync);
2305 p->state = TASK_RUNNING;
2307 if (p->sched_class->task_wake_up)
2308 p->sched_class->task_wake_up(rq, p);
2311 current->se.last_wakeup = current->se.sum_exec_runtime;
2313 task_rq_unlock(rq, &flags);
2318 int wake_up_process(struct task_struct *p)
2320 return try_to_wake_up(p, TASK_ALL, 0);
2322 EXPORT_SYMBOL(wake_up_process);
2324 int wake_up_state(struct task_struct *p, unsigned int state)
2326 return try_to_wake_up(p, state, 0);
2330 * Perform scheduler related setup for a newly forked process p.
2331 * p is forked by current.
2333 * __sched_fork() is basic setup used by init_idle() too:
2335 static void __sched_fork(struct task_struct *p)
2337 p->se.exec_start = 0;
2338 p->se.sum_exec_runtime = 0;
2339 p->se.prev_sum_exec_runtime = 0;
2340 p->se.last_wakeup = 0;
2341 p->se.avg_overlap = 0;
2343 #ifdef CONFIG_SCHEDSTATS
2344 p->se.wait_start = 0;
2345 p->se.sum_sleep_runtime = 0;
2346 p->se.sleep_start = 0;
2347 p->se.block_start = 0;
2348 p->se.sleep_max = 0;
2349 p->se.block_max = 0;
2351 p->se.slice_max = 0;
2355 INIT_LIST_HEAD(&p->rt.run_list);
2357 INIT_LIST_HEAD(&p->se.group_node);
2359 #ifdef CONFIG_PREEMPT_NOTIFIERS
2360 INIT_HLIST_HEAD(&p->preempt_notifiers);
2364 * We mark the process as running here, but have not actually
2365 * inserted it onto the runqueue yet. This guarantees that
2366 * nobody will actually run it, and a signal or other external
2367 * event cannot wake it up and insert it on the runqueue either.
2369 p->state = TASK_RUNNING;
2373 * fork()/clone()-time setup:
2375 void sched_fork(struct task_struct *p, int clone_flags)
2377 int cpu = get_cpu();
2382 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2384 set_task_cpu(p, cpu);
2387 * Make sure we do not leak PI boosting priority to the child:
2389 p->prio = current->normal_prio;
2390 if (!rt_prio(p->prio))
2391 p->sched_class = &fair_sched_class;
2393 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2394 if (likely(sched_info_on()))
2395 memset(&p->sched_info, 0, sizeof(p->sched_info));
2397 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2400 #ifdef CONFIG_PREEMPT
2401 /* Want to start with kernel preemption disabled. */
2402 task_thread_info(p)->preempt_count = 1;
2408 * wake_up_new_task - wake up a newly created task for the first time.
2410 * This function will do some initial scheduler statistics housekeeping
2411 * that must be done for every newly created context, then puts the task
2412 * on the runqueue and wakes it.
2414 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2416 unsigned long flags;
2419 rq = task_rq_lock(p, &flags);
2420 BUG_ON(p->state != TASK_RUNNING);
2421 update_rq_clock(rq);
2423 p->prio = effective_prio(p);
2425 if (!p->sched_class->task_new || !current->se.on_rq) {
2426 activate_task(rq, p, 0);
2429 * Let the scheduling class do new task startup
2430 * management (if any):
2432 p->sched_class->task_new(rq, p);
2435 trace_mark(kernel_sched_wakeup_new,
2436 "pid %d state %ld ## rq %p task %p rq->curr %p",
2437 p->pid, p->state, rq, p, rq->curr);
2438 check_preempt_curr(rq, p, 0);
2440 if (p->sched_class->task_wake_up)
2441 p->sched_class->task_wake_up(rq, p);
2443 task_rq_unlock(rq, &flags);
2446 #ifdef CONFIG_PREEMPT_NOTIFIERS
2449 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2450 * @notifier: notifier struct to register
2452 void preempt_notifier_register(struct preempt_notifier *notifier)
2454 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2456 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2459 * preempt_notifier_unregister - no longer interested in preemption notifications
2460 * @notifier: notifier struct to unregister
2462 * This is safe to call from within a preemption notifier.
2464 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2466 hlist_del(¬ifier->link);
2468 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2470 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2472 struct preempt_notifier *notifier;
2473 struct hlist_node *node;
2475 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2476 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2480 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2481 struct task_struct *next)
2483 struct preempt_notifier *notifier;
2484 struct hlist_node *node;
2486 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2487 notifier->ops->sched_out(notifier, next);
2490 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2492 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2497 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2498 struct task_struct *next)
2502 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2505 * prepare_task_switch - prepare to switch tasks
2506 * @rq: the runqueue preparing to switch
2507 * @prev: the current task that is being switched out
2508 * @next: the task we are going to switch to.
2510 * This is called with the rq lock held and interrupts off. It must
2511 * be paired with a subsequent finish_task_switch after the context
2514 * prepare_task_switch sets up locking and calls architecture specific
2518 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2519 struct task_struct *next)
2521 fire_sched_out_preempt_notifiers(prev, next);
2522 prepare_lock_switch(rq, next);
2523 prepare_arch_switch(next);
2527 * finish_task_switch - clean up after a task-switch
2528 * @rq: runqueue associated with task-switch
2529 * @prev: the thread we just switched away from.
2531 * finish_task_switch must be called after the context switch, paired
2532 * with a prepare_task_switch call before the context switch.
2533 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2534 * and do any other architecture-specific cleanup actions.
2536 * Note that we may have delayed dropping an mm in context_switch(). If
2537 * so, we finish that here outside of the runqueue lock. (Doing it
2538 * with the lock held can cause deadlocks; see schedule() for
2541 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2542 __releases(rq->lock)
2544 struct mm_struct *mm = rq->prev_mm;
2550 * A task struct has one reference for the use as "current".
2551 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2552 * schedule one last time. The schedule call will never return, and
2553 * the scheduled task must drop that reference.
2554 * The test for TASK_DEAD must occur while the runqueue locks are
2555 * still held, otherwise prev could be scheduled on another cpu, die
2556 * there before we look at prev->state, and then the reference would
2558 * Manfred Spraul <manfred@colorfullife.com>
2560 prev_state = prev->state;
2561 finish_arch_switch(prev);
2562 finish_lock_switch(rq, prev);
2564 if (current->sched_class->post_schedule)
2565 current->sched_class->post_schedule(rq);
2568 fire_sched_in_preempt_notifiers(current);
2571 if (unlikely(prev_state == TASK_DEAD)) {
2573 * Remove function-return probe instances associated with this
2574 * task and put them back on the free list.
2576 kprobe_flush_task(prev);
2577 put_task_struct(prev);
2582 * schedule_tail - first thing a freshly forked thread must call.
2583 * @prev: the thread we just switched away from.
2585 asmlinkage void schedule_tail(struct task_struct *prev)
2586 __releases(rq->lock)
2588 struct rq *rq = this_rq();
2590 finish_task_switch(rq, prev);
2591 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2592 /* In this case, finish_task_switch does not reenable preemption */
2595 if (current->set_child_tid)
2596 put_user(task_pid_vnr(current), current->set_child_tid);
2600 * context_switch - switch to the new MM and the new
2601 * thread's register state.
2604 context_switch(struct rq *rq, struct task_struct *prev,
2605 struct task_struct *next)
2607 struct mm_struct *mm, *oldmm;
2609 prepare_task_switch(rq, prev, next);
2610 trace_mark(kernel_sched_schedule,
2611 "prev_pid %d next_pid %d prev_state %ld "
2612 "## rq %p prev %p next %p",
2613 prev->pid, next->pid, prev->state,
2616 oldmm = prev->active_mm;
2618 * For paravirt, this is coupled with an exit in switch_to to
2619 * combine the page table reload and the switch backend into
2622 arch_enter_lazy_cpu_mode();
2624 if (unlikely(!mm)) {
2625 next->active_mm = oldmm;
2626 atomic_inc(&oldmm->mm_count);
2627 enter_lazy_tlb(oldmm, next);
2629 switch_mm(oldmm, mm, next);
2631 if (unlikely(!prev->mm)) {
2632 prev->active_mm = NULL;
2633 rq->prev_mm = oldmm;
2636 * Since the runqueue lock will be released by the next
2637 * task (which is an invalid locking op but in the case
2638 * of the scheduler it's an obvious special-case), so we
2639 * do an early lockdep release here:
2641 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2642 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2645 /* Here we just switch the register state and the stack. */
2646 switch_to(prev, next, prev);
2650 * this_rq must be evaluated again because prev may have moved
2651 * CPUs since it called schedule(), thus the 'rq' on its stack
2652 * frame will be invalid.
2654 finish_task_switch(this_rq(), prev);
2658 * nr_running, nr_uninterruptible and nr_context_switches:
2660 * externally visible scheduler statistics: current number of runnable
2661 * threads, current number of uninterruptible-sleeping threads, total
2662 * number of context switches performed since bootup.
2664 unsigned long nr_running(void)
2666 unsigned long i, sum = 0;
2668 for_each_online_cpu(i)
2669 sum += cpu_rq(i)->nr_running;
2674 unsigned long nr_uninterruptible(void)
2676 unsigned long i, sum = 0;
2678 for_each_possible_cpu(i)
2679 sum += cpu_rq(i)->nr_uninterruptible;
2682 * Since we read the counters lockless, it might be slightly
2683 * inaccurate. Do not allow it to go below zero though:
2685 if (unlikely((long)sum < 0))
2691 unsigned long long nr_context_switches(void)
2694 unsigned long long sum = 0;
2696 for_each_possible_cpu(i)
2697 sum += cpu_rq(i)->nr_switches;
2702 unsigned long nr_iowait(void)
2704 unsigned long i, sum = 0;
2706 for_each_possible_cpu(i)
2707 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2712 unsigned long nr_active(void)
2714 unsigned long i, running = 0, uninterruptible = 0;
2716 for_each_online_cpu(i) {
2717 running += cpu_rq(i)->nr_running;
2718 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2721 if (unlikely((long)uninterruptible < 0))
2722 uninterruptible = 0;
2724 return running + uninterruptible;
2728 * Update rq->cpu_load[] statistics. This function is usually called every
2729 * scheduler tick (TICK_NSEC).
2731 static void update_cpu_load(struct rq *this_rq)
2733 unsigned long this_load = this_rq->load.weight;
2736 this_rq->nr_load_updates++;
2738 /* Update our load: */
2739 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2740 unsigned long old_load, new_load;
2742 /* scale is effectively 1 << i now, and >> i divides by scale */
2744 old_load = this_rq->cpu_load[i];
2745 new_load = this_load;
2747 * Round up the averaging division if load is increasing. This
2748 * prevents us from getting stuck on 9 if the load is 10, for
2751 if (new_load > old_load)
2752 new_load += scale-1;
2753 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2760 * double_rq_lock - safely lock two runqueues
2762 * Note this does not disable interrupts like task_rq_lock,
2763 * you need to do so manually before calling.
2765 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2766 __acquires(rq1->lock)
2767 __acquires(rq2->lock)
2769 BUG_ON(!irqs_disabled());
2771 spin_lock(&rq1->lock);
2772 __acquire(rq2->lock); /* Fake it out ;) */
2775 spin_lock(&rq1->lock);
2776 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2778 spin_lock(&rq2->lock);
2779 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2782 update_rq_clock(rq1);
2783 update_rq_clock(rq2);
2787 * double_rq_unlock - safely unlock two runqueues
2789 * Note this does not restore interrupts like task_rq_unlock,
2790 * you need to do so manually after calling.
2792 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2793 __releases(rq1->lock)
2794 __releases(rq2->lock)
2796 spin_unlock(&rq1->lock);
2798 spin_unlock(&rq2->lock);
2800 __release(rq2->lock);
2804 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2806 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2807 __releases(this_rq->lock)
2808 __acquires(busiest->lock)
2809 __acquires(this_rq->lock)
2813 if (unlikely(!irqs_disabled())) {
2814 /* printk() doesn't work good under rq->lock */
2815 spin_unlock(&this_rq->lock);
2818 if (unlikely(!spin_trylock(&busiest->lock))) {
2819 if (busiest < this_rq) {
2820 spin_unlock(&this_rq->lock);
2821 spin_lock(&busiest->lock);
2822 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
2825 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
2830 static void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2831 __releases(busiest->lock)
2833 spin_unlock(&busiest->lock);
2834 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
2838 * If dest_cpu is allowed for this process, migrate the task to it.
2839 * This is accomplished by forcing the cpu_allowed mask to only
2840 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2841 * the cpu_allowed mask is restored.
2843 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2845 struct migration_req req;
2846 unsigned long flags;
2849 rq = task_rq_lock(p, &flags);
2850 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2851 || unlikely(!cpu_active(dest_cpu)))
2854 /* force the process onto the specified CPU */
2855 if (migrate_task(p, dest_cpu, &req)) {
2856 /* Need to wait for migration thread (might exit: take ref). */
2857 struct task_struct *mt = rq->migration_thread;
2859 get_task_struct(mt);
2860 task_rq_unlock(rq, &flags);
2861 wake_up_process(mt);
2862 put_task_struct(mt);
2863 wait_for_completion(&req.done);
2868 task_rq_unlock(rq, &flags);
2872 * sched_exec - execve() is a valuable balancing opportunity, because at
2873 * this point the task has the smallest effective memory and cache footprint.
2875 void sched_exec(void)
2877 int new_cpu, this_cpu = get_cpu();
2878 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2880 if (new_cpu != this_cpu)
2881 sched_migrate_task(current, new_cpu);
2885 * pull_task - move a task from a remote runqueue to the local runqueue.
2886 * Both runqueues must be locked.
2888 static void pull_task(struct rq *src_rq, struct task_struct *p,
2889 struct rq *this_rq, int this_cpu)
2891 deactivate_task(src_rq, p, 0);
2892 set_task_cpu(p, this_cpu);
2893 activate_task(this_rq, p, 0);
2895 * Note that idle threads have a prio of MAX_PRIO, for this test
2896 * to be always true for them.
2898 check_preempt_curr(this_rq, p, 0);
2902 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2905 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2906 struct sched_domain *sd, enum cpu_idle_type idle,
2910 * We do not migrate tasks that are:
2911 * 1) running (obviously), or
2912 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2913 * 3) are cache-hot on their current CPU.
2915 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2916 schedstat_inc(p, se.nr_failed_migrations_affine);
2921 if (task_running(rq, p)) {
2922 schedstat_inc(p, se.nr_failed_migrations_running);
2927 * Aggressive migration if:
2928 * 1) task is cache cold, or
2929 * 2) too many balance attempts have failed.
2932 if (!task_hot(p, rq->clock, sd) ||
2933 sd->nr_balance_failed > sd->cache_nice_tries) {
2934 #ifdef CONFIG_SCHEDSTATS
2935 if (task_hot(p, rq->clock, sd)) {
2936 schedstat_inc(sd, lb_hot_gained[idle]);
2937 schedstat_inc(p, se.nr_forced_migrations);
2943 if (task_hot(p, rq->clock, sd)) {
2944 schedstat_inc(p, se.nr_failed_migrations_hot);
2950 static unsigned long
2951 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2952 unsigned long max_load_move, struct sched_domain *sd,
2953 enum cpu_idle_type idle, int *all_pinned,
2954 int *this_best_prio, struct rq_iterator *iterator)
2956 int loops = 0, pulled = 0, pinned = 0;
2957 struct task_struct *p;
2958 long rem_load_move = max_load_move;
2960 if (max_load_move == 0)
2966 * Start the load-balancing iterator:
2968 p = iterator->start(iterator->arg);
2970 if (!p || loops++ > sysctl_sched_nr_migrate)
2973 if ((p->se.load.weight >> 1) > rem_load_move ||
2974 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2975 p = iterator->next(iterator->arg);
2979 pull_task(busiest, p, this_rq, this_cpu);
2981 rem_load_move -= p->se.load.weight;
2984 * We only want to steal up to the prescribed amount of weighted load.
2986 if (rem_load_move > 0) {
2987 if (p->prio < *this_best_prio)
2988 *this_best_prio = p->prio;
2989 p = iterator->next(iterator->arg);
2994 * Right now, this is one of only two places pull_task() is called,
2995 * so we can safely collect pull_task() stats here rather than
2996 * inside pull_task().
2998 schedstat_add(sd, lb_gained[idle], pulled);
3001 *all_pinned = pinned;
3003 return max_load_move - rem_load_move;
3007 * move_tasks tries to move up to max_load_move weighted load from busiest to
3008 * this_rq, as part of a balancing operation within domain "sd".
3009 * Returns 1 if successful and 0 otherwise.
3011 * Called with both runqueues locked.
3013 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3014 unsigned long max_load_move,
3015 struct sched_domain *sd, enum cpu_idle_type idle,
3018 const struct sched_class *class = sched_class_highest;
3019 unsigned long total_load_moved = 0;
3020 int this_best_prio = this_rq->curr->prio;
3024 class->load_balance(this_rq, this_cpu, busiest,
3025 max_load_move - total_load_moved,
3026 sd, idle, all_pinned, &this_best_prio);
3027 class = class->next;
3029 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3032 } while (class && max_load_move > total_load_moved);
3034 return total_load_moved > 0;
3038 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3039 struct sched_domain *sd, enum cpu_idle_type idle,
3040 struct rq_iterator *iterator)
3042 struct task_struct *p = iterator->start(iterator->arg);
3046 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3047 pull_task(busiest, p, this_rq, this_cpu);
3049 * Right now, this is only the second place pull_task()
3050 * is called, so we can safely collect pull_task()
3051 * stats here rather than inside pull_task().
3053 schedstat_inc(sd, lb_gained[idle]);
3057 p = iterator->next(iterator->arg);
3064 * move_one_task tries to move exactly one task from busiest to this_rq, as
3065 * part of active balancing operations within "domain".
3066 * Returns 1 if successful and 0 otherwise.
3068 * Called with both runqueues locked.
3070 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3071 struct sched_domain *sd, enum cpu_idle_type idle)
3073 const struct sched_class *class;
3075 for (class = sched_class_highest; class; class = class->next)
3076 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3083 * find_busiest_group finds and returns the busiest CPU group within the
3084 * domain. It calculates and returns the amount of weighted load which
3085 * should be moved to restore balance via the imbalance parameter.
3087 static struct sched_group *
3088 find_busiest_group(struct sched_domain *sd, int this_cpu,
3089 unsigned long *imbalance, enum cpu_idle_type idle,
3090 int *sd_idle, const cpumask_t *cpus, int *balance)
3092 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3093 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3094 unsigned long max_pull;
3095 unsigned long busiest_load_per_task, busiest_nr_running;
3096 unsigned long this_load_per_task, this_nr_running;
3097 int load_idx, group_imb = 0;
3098 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3099 int power_savings_balance = 1;
3100 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3101 unsigned long min_nr_running = ULONG_MAX;
3102 struct sched_group *group_min = NULL, *group_leader = NULL;
3105 max_load = this_load = total_load = total_pwr = 0;
3106 busiest_load_per_task = busiest_nr_running = 0;
3107 this_load_per_task = this_nr_running = 0;
3109 if (idle == CPU_NOT_IDLE)
3110 load_idx = sd->busy_idx;
3111 else if (idle == CPU_NEWLY_IDLE)
3112 load_idx = sd->newidle_idx;
3114 load_idx = sd->idle_idx;
3117 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3120 int __group_imb = 0;
3121 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3122 unsigned long sum_nr_running, sum_weighted_load;
3123 unsigned long sum_avg_load_per_task;
3124 unsigned long avg_load_per_task;
3126 local_group = cpu_isset(this_cpu, group->cpumask);
3129 balance_cpu = first_cpu(group->cpumask);
3131 /* Tally up the load of all CPUs in the group */
3132 sum_weighted_load = sum_nr_running = avg_load = 0;
3133 sum_avg_load_per_task = avg_load_per_task = 0;
3136 min_cpu_load = ~0UL;
3138 for_each_cpu_mask_nr(i, group->cpumask) {
3141 if (!cpu_isset(i, *cpus))
3146 if (*sd_idle && rq->nr_running)
3149 /* Bias balancing toward cpus of our domain */
3151 if (idle_cpu(i) && !first_idle_cpu) {
3156 load = target_load(i, load_idx);
3158 load = source_load(i, load_idx);
3159 if (load > max_cpu_load)
3160 max_cpu_load = load;
3161 if (min_cpu_load > load)
3162 min_cpu_load = load;
3166 sum_nr_running += rq->nr_running;
3167 sum_weighted_load += weighted_cpuload(i);
3169 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3173 * First idle cpu or the first cpu(busiest) in this sched group
3174 * is eligible for doing load balancing at this and above
3175 * domains. In the newly idle case, we will allow all the cpu's
3176 * to do the newly idle load balance.
3178 if (idle != CPU_NEWLY_IDLE && local_group &&
3179 balance_cpu != this_cpu && balance) {
3184 total_load += avg_load;
3185 total_pwr += group->__cpu_power;
3187 /* Adjust by relative CPU power of the group */
3188 avg_load = sg_div_cpu_power(group,
3189 avg_load * SCHED_LOAD_SCALE);
3193 * Consider the group unbalanced when the imbalance is larger
3194 * than the average weight of two tasks.
3196 * APZ: with cgroup the avg task weight can vary wildly and
3197 * might not be a suitable number - should we keep a
3198 * normalized nr_running number somewhere that negates
3201 avg_load_per_task = sg_div_cpu_power(group,
3202 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3204 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3207 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3210 this_load = avg_load;
3212 this_nr_running = sum_nr_running;
3213 this_load_per_task = sum_weighted_load;
3214 } else if (avg_load > max_load &&
3215 (sum_nr_running > group_capacity || __group_imb)) {
3216 max_load = avg_load;
3218 busiest_nr_running = sum_nr_running;
3219 busiest_load_per_task = sum_weighted_load;
3220 group_imb = __group_imb;
3223 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3225 * Busy processors will not participate in power savings
3228 if (idle == CPU_NOT_IDLE ||
3229 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3233 * If the local group is idle or completely loaded
3234 * no need to do power savings balance at this domain
3236 if (local_group && (this_nr_running >= group_capacity ||
3238 power_savings_balance = 0;
3241 * If a group is already running at full capacity or idle,
3242 * don't include that group in power savings calculations
3244 if (!power_savings_balance || sum_nr_running >= group_capacity
3249 * Calculate the group which has the least non-idle load.
3250 * This is the group from where we need to pick up the load
3253 if ((sum_nr_running < min_nr_running) ||
3254 (sum_nr_running == min_nr_running &&
3255 first_cpu(group->cpumask) <
3256 first_cpu(group_min->cpumask))) {
3258 min_nr_running = sum_nr_running;
3259 min_load_per_task = sum_weighted_load /
3264 * Calculate the group which is almost near its
3265 * capacity but still has some space to pick up some load
3266 * from other group and save more power
3268 if (sum_nr_running <= group_capacity - 1) {
3269 if (sum_nr_running > leader_nr_running ||
3270 (sum_nr_running == leader_nr_running &&
3271 first_cpu(group->cpumask) >
3272 first_cpu(group_leader->cpumask))) {
3273 group_leader = group;
3274 leader_nr_running = sum_nr_running;
3279 group = group->next;
3280 } while (group != sd->groups);
3282 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3285 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3287 if (this_load >= avg_load ||
3288 100*max_load <= sd->imbalance_pct*this_load)
3291 busiest_load_per_task /= busiest_nr_running;
3293 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3296 * We're trying to get all the cpus to the average_load, so we don't
3297 * want to push ourselves above the average load, nor do we wish to
3298 * reduce the max loaded cpu below the average load, as either of these
3299 * actions would just result in more rebalancing later, and ping-pong
3300 * tasks around. Thus we look for the minimum possible imbalance.
3301 * Negative imbalances (*we* are more loaded than anyone else) will
3302 * be counted as no imbalance for these purposes -- we can't fix that
3303 * by pulling tasks to us. Be careful of negative numbers as they'll
3304 * appear as very large values with unsigned longs.
3306 if (max_load <= busiest_load_per_task)
3310 * In the presence of smp nice balancing, certain scenarios can have
3311 * max load less than avg load(as we skip the groups at or below
3312 * its cpu_power, while calculating max_load..)
3314 if (max_load < avg_load) {
3316 goto small_imbalance;
3319 /* Don't want to pull so many tasks that a group would go idle */
3320 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3322 /* How much load to actually move to equalise the imbalance */
3323 *imbalance = min(max_pull * busiest->__cpu_power,
3324 (avg_load - this_load) * this->__cpu_power)
3328 * if *imbalance is less than the average load per runnable task
3329 * there is no gaurantee that any tasks will be moved so we'll have
3330 * a think about bumping its value to force at least one task to be
3333 if (*imbalance < busiest_load_per_task) {
3334 unsigned long tmp, pwr_now, pwr_move;
3338 pwr_move = pwr_now = 0;
3340 if (this_nr_running) {
3341 this_load_per_task /= this_nr_running;
3342 if (busiest_load_per_task > this_load_per_task)
3345 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3347 if (max_load - this_load + 2*busiest_load_per_task >=
3348 busiest_load_per_task * imbn) {
3349 *imbalance = busiest_load_per_task;
3354 * OK, we don't have enough imbalance to justify moving tasks,
3355 * however we may be able to increase total CPU power used by
3359 pwr_now += busiest->__cpu_power *
3360 min(busiest_load_per_task, max_load);
3361 pwr_now += this->__cpu_power *
3362 min(this_load_per_task, this_load);
3363 pwr_now /= SCHED_LOAD_SCALE;
3365 /* Amount of load we'd subtract */
3366 tmp = sg_div_cpu_power(busiest,
3367 busiest_load_per_task * SCHED_LOAD_SCALE);
3369 pwr_move += busiest->__cpu_power *
3370 min(busiest_load_per_task, max_load - tmp);
3372 /* Amount of load we'd add */
3373 if (max_load * busiest->__cpu_power <
3374 busiest_load_per_task * SCHED_LOAD_SCALE)
3375 tmp = sg_div_cpu_power(this,
3376 max_load * busiest->__cpu_power);
3378 tmp = sg_div_cpu_power(this,
3379 busiest_load_per_task * SCHED_LOAD_SCALE);
3380 pwr_move += this->__cpu_power *
3381 min(this_load_per_task, this_load + tmp);
3382 pwr_move /= SCHED_LOAD_SCALE;
3384 /* Move if we gain throughput */
3385 if (pwr_move > pwr_now)
3386 *imbalance = busiest_load_per_task;
3392 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3393 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3396 if (this == group_leader && group_leader != group_min) {
3397 *imbalance = min_load_per_task;
3407 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3410 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3411 unsigned long imbalance, const cpumask_t *cpus)
3413 struct rq *busiest = NULL, *rq;
3414 unsigned long max_load = 0;
3417 for_each_cpu_mask_nr(i, group->cpumask) {
3420 if (!cpu_isset(i, *cpus))
3424 wl = weighted_cpuload(i);
3426 if (rq->nr_running == 1 && wl > imbalance)
3429 if (wl > max_load) {
3439 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3440 * so long as it is large enough.
3442 #define MAX_PINNED_INTERVAL 512
3445 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3446 * tasks if there is an imbalance.
3448 static int load_balance(int this_cpu, struct rq *this_rq,
3449 struct sched_domain *sd, enum cpu_idle_type idle,
3450 int *balance, cpumask_t *cpus)
3452 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3453 struct sched_group *group;
3454 unsigned long imbalance;
3456 unsigned long flags;
3461 * When power savings policy is enabled for the parent domain, idle
3462 * sibling can pick up load irrespective of busy siblings. In this case,
3463 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3464 * portraying it as CPU_NOT_IDLE.
3466 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3467 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3470 schedstat_inc(sd, lb_count[idle]);
3474 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3481 schedstat_inc(sd, lb_nobusyg[idle]);
3485 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3487 schedstat_inc(sd, lb_nobusyq[idle]);
3491 BUG_ON(busiest == this_rq);
3493 schedstat_add(sd, lb_imbalance[idle], imbalance);
3496 if (busiest->nr_running > 1) {
3498 * Attempt to move tasks. If find_busiest_group has found
3499 * an imbalance but busiest->nr_running <= 1, the group is
3500 * still unbalanced. ld_moved simply stays zero, so it is
3501 * correctly treated as an imbalance.
3503 local_irq_save(flags);
3504 double_rq_lock(this_rq, busiest);
3505 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3506 imbalance, sd, idle, &all_pinned);
3507 double_rq_unlock(this_rq, busiest);
3508 local_irq_restore(flags);
3511 * some other cpu did the load balance for us.
3513 if (ld_moved && this_cpu != smp_processor_id())
3514 resched_cpu(this_cpu);
3516 /* All tasks on this runqueue were pinned by CPU affinity */
3517 if (unlikely(all_pinned)) {
3518 cpu_clear(cpu_of(busiest), *cpus);
3519 if (!cpus_empty(*cpus))
3526 schedstat_inc(sd, lb_failed[idle]);
3527 sd->nr_balance_failed++;
3529 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3531 spin_lock_irqsave(&busiest->lock, flags);
3533 /* don't kick the migration_thread, if the curr
3534 * task on busiest cpu can't be moved to this_cpu
3536 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3537 spin_unlock_irqrestore(&busiest->lock, flags);
3539 goto out_one_pinned;
3542 if (!busiest->active_balance) {
3543 busiest->active_balance = 1;
3544 busiest->push_cpu = this_cpu;
3547 spin_unlock_irqrestore(&busiest->lock, flags);
3549 wake_up_process(busiest->migration_thread);
3552 * We've kicked active balancing, reset the failure
3555 sd->nr_balance_failed = sd->cache_nice_tries+1;
3558 sd->nr_balance_failed = 0;
3560 if (likely(!active_balance)) {
3561 /* We were unbalanced, so reset the balancing interval */
3562 sd->balance_interval = sd->min_interval;
3565 * If we've begun active balancing, start to back off. This
3566 * case may not be covered by the all_pinned logic if there
3567 * is only 1 task on the busy runqueue (because we don't call
3570 if (sd->balance_interval < sd->max_interval)
3571 sd->balance_interval *= 2;
3574 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3575 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3581 schedstat_inc(sd, lb_balanced[idle]);
3583 sd->nr_balance_failed = 0;
3586 /* tune up the balancing interval */
3587 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3588 (sd->balance_interval < sd->max_interval))
3589 sd->balance_interval *= 2;
3591 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3592 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3603 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3604 * tasks if there is an imbalance.
3606 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3607 * this_rq is locked.
3610 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3613 struct sched_group *group;
3614 struct rq *busiest = NULL;
3615 unsigned long imbalance;
3623 * When power savings policy is enabled for the parent domain, idle
3624 * sibling can pick up load irrespective of busy siblings. In this case,
3625 * let the state of idle sibling percolate up as IDLE, instead of
3626 * portraying it as CPU_NOT_IDLE.
3628 if (sd->flags & SD_SHARE_CPUPOWER &&
3629 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3632 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3634 update_shares_locked(this_rq, sd);
3635 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3636 &sd_idle, cpus, NULL);
3638 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3642 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3644 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3648 BUG_ON(busiest == this_rq);
3650 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3653 if (busiest->nr_running > 1) {
3654 /* Attempt to move tasks */
3655 double_lock_balance(this_rq, busiest);
3656 /* this_rq->clock is already updated */
3657 update_rq_clock(busiest);
3658 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3659 imbalance, sd, CPU_NEWLY_IDLE,
3661 double_unlock_balance(this_rq, busiest);
3663 if (unlikely(all_pinned)) {
3664 cpu_clear(cpu_of(busiest), *cpus);
3665 if (!cpus_empty(*cpus))
3671 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3672 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3673 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3676 sd->nr_balance_failed = 0;
3678 update_shares_locked(this_rq, sd);
3682 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3683 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3684 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3686 sd->nr_balance_failed = 0;
3692 * idle_balance is called by schedule() if this_cpu is about to become
3693 * idle. Attempts to pull tasks from other CPUs.
3695 static void idle_balance(int this_cpu, struct rq *this_rq)
3697 struct sched_domain *sd;
3698 int pulled_task = -1;
3699 unsigned long next_balance = jiffies + HZ;
3702 for_each_domain(this_cpu, sd) {
3703 unsigned long interval;
3705 if (!(sd->flags & SD_LOAD_BALANCE))
3708 if (sd->flags & SD_BALANCE_NEWIDLE)
3709 /* If we've pulled tasks over stop searching: */
3710 pulled_task = load_balance_newidle(this_cpu, this_rq,
3713 interval = msecs_to_jiffies(sd->balance_interval);
3714 if (time_after(next_balance, sd->last_balance + interval))
3715 next_balance = sd->last_balance + interval;
3719 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3721 * We are going idle. next_balance may be set based on
3722 * a busy processor. So reset next_balance.
3724 this_rq->next_balance = next_balance;
3729 * active_load_balance is run by migration threads. It pushes running tasks
3730 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3731 * running on each physical CPU where possible, and avoids physical /
3732 * logical imbalances.
3734 * Called with busiest_rq locked.
3736 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3738 int target_cpu = busiest_rq->push_cpu;
3739 struct sched_domain *sd;
3740 struct rq *target_rq;
3742 /* Is there any task to move? */
3743 if (busiest_rq->nr_running <= 1)
3746 target_rq = cpu_rq(target_cpu);
3749 * This condition is "impossible", if it occurs
3750 * we need to fix it. Originally reported by
3751 * Bjorn Helgaas on a 128-cpu setup.
3753 BUG_ON(busiest_rq == target_rq);
3755 /* move a task from busiest_rq to target_rq */
3756 double_lock_balance(busiest_rq, target_rq);
3757 update_rq_clock(busiest_rq);
3758 update_rq_clock(target_rq);
3760 /* Search for an sd spanning us and the target CPU. */
3761 for_each_domain(target_cpu, sd) {
3762 if ((sd->flags & SD_LOAD_BALANCE) &&
3763 cpu_isset(busiest_cpu, sd->span))
3768 schedstat_inc(sd, alb_count);
3770 if (move_one_task(target_rq, target_cpu, busiest_rq,
3772 schedstat_inc(sd, alb_pushed);
3774 schedstat_inc(sd, alb_failed);
3776 double_unlock_balance(busiest_rq, target_rq);
3781 atomic_t load_balancer;
3783 } nohz ____cacheline_aligned = {
3784 .load_balancer = ATOMIC_INIT(-1),
3785 .cpu_mask = CPU_MASK_NONE,
3789 * This routine will try to nominate the ilb (idle load balancing)
3790 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3791 * load balancing on behalf of all those cpus. If all the cpus in the system
3792 * go into this tickless mode, then there will be no ilb owner (as there is
3793 * no need for one) and all the cpus will sleep till the next wakeup event
3796 * For the ilb owner, tick is not stopped. And this tick will be used
3797 * for idle load balancing. ilb owner will still be part of
3800 * While stopping the tick, this cpu will become the ilb owner if there
3801 * is no other owner. And will be the owner till that cpu becomes busy
3802 * or if all cpus in the system stop their ticks at which point
3803 * there is no need for ilb owner.
3805 * When the ilb owner becomes busy, it nominates another owner, during the
3806 * next busy scheduler_tick()
3808 int select_nohz_load_balancer(int stop_tick)
3810 int cpu = smp_processor_id();
3813 cpu_set(cpu, nohz.cpu_mask);
3814 cpu_rq(cpu)->in_nohz_recently = 1;
3817 * If we are going offline and still the leader, give up!
3819 if (!cpu_active(cpu) &&
3820 atomic_read(&nohz.load_balancer) == cpu) {
3821 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3826 /* time for ilb owner also to sleep */
3827 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3828 if (atomic_read(&nohz.load_balancer) == cpu)
3829 atomic_set(&nohz.load_balancer, -1);
3833 if (atomic_read(&nohz.load_balancer) == -1) {
3834 /* make me the ilb owner */
3835 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3837 } else if (atomic_read(&nohz.load_balancer) == cpu)
3840 if (!cpu_isset(cpu, nohz.cpu_mask))
3843 cpu_clear(cpu, nohz.cpu_mask);
3845 if (atomic_read(&nohz.load_balancer) == cpu)
3846 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3853 static DEFINE_SPINLOCK(balancing);
3856 * It checks each scheduling domain to see if it is due to be balanced,
3857 * and initiates a balancing operation if so.
3859 * Balancing parameters are set up in arch_init_sched_domains.
3861 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3864 struct rq *rq = cpu_rq(cpu);
3865 unsigned long interval;
3866 struct sched_domain *sd;
3867 /* Earliest time when we have to do rebalance again */
3868 unsigned long next_balance = jiffies + 60*HZ;
3869 int update_next_balance = 0;
3873 for_each_domain(cpu, sd) {
3874 if (!(sd->flags & SD_LOAD_BALANCE))
3877 interval = sd->balance_interval;
3878 if (idle != CPU_IDLE)
3879 interval *= sd->busy_factor;
3881 /* scale ms to jiffies */
3882 interval = msecs_to_jiffies(interval);
3883 if (unlikely(!interval))
3885 if (interval > HZ*NR_CPUS/10)
3886 interval = HZ*NR_CPUS/10;
3888 need_serialize = sd->flags & SD_SERIALIZE;
3890 if (need_serialize) {
3891 if (!spin_trylock(&balancing))
3895 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3896 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3898 * We've pulled tasks over so either we're no
3899 * longer idle, or one of our SMT siblings is
3902 idle = CPU_NOT_IDLE;
3904 sd->last_balance = jiffies;
3907 spin_unlock(&balancing);
3909 if (time_after(next_balance, sd->last_balance + interval)) {
3910 next_balance = sd->last_balance + interval;
3911 update_next_balance = 1;
3915 * Stop the load balance at this level. There is another
3916 * CPU in our sched group which is doing load balancing more
3924 * next_balance will be updated only when there is a need.
3925 * When the cpu is attached to null domain for ex, it will not be
3928 if (likely(update_next_balance))
3929 rq->next_balance = next_balance;
3933 * run_rebalance_domains is triggered when needed from the scheduler tick.
3934 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3935 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3937 static void run_rebalance_domains(struct softirq_action *h)
3939 int this_cpu = smp_processor_id();
3940 struct rq *this_rq = cpu_rq(this_cpu);
3941 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3942 CPU_IDLE : CPU_NOT_IDLE;
3944 rebalance_domains(this_cpu, idle);
3948 * If this cpu is the owner for idle load balancing, then do the
3949 * balancing on behalf of the other idle cpus whose ticks are
3952 if (this_rq->idle_at_tick &&
3953 atomic_read(&nohz.load_balancer) == this_cpu) {
3954 cpumask_t cpus = nohz.cpu_mask;
3958 cpu_clear(this_cpu, cpus);
3959 for_each_cpu_mask_nr(balance_cpu, cpus) {
3961 * If this cpu gets work to do, stop the load balancing
3962 * work being done for other cpus. Next load
3963 * balancing owner will pick it up.
3968 rebalance_domains(balance_cpu, CPU_IDLE);
3970 rq = cpu_rq(balance_cpu);
3971 if (time_after(this_rq->next_balance, rq->next_balance))
3972 this_rq->next_balance = rq->next_balance;
3979 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3981 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3982 * idle load balancing owner or decide to stop the periodic load balancing,
3983 * if the whole system is idle.
3985 static inline void trigger_load_balance(struct rq *rq, int cpu)
3989 * If we were in the nohz mode recently and busy at the current
3990 * scheduler tick, then check if we need to nominate new idle
3993 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3994 rq->in_nohz_recently = 0;
3996 if (atomic_read(&nohz.load_balancer) == cpu) {
3997 cpu_clear(cpu, nohz.cpu_mask);
3998 atomic_set(&nohz.load_balancer, -1);
4001 if (atomic_read(&nohz.load_balancer) == -1) {
4003 * simple selection for now: Nominate the
4004 * first cpu in the nohz list to be the next
4007 * TBD: Traverse the sched domains and nominate
4008 * the nearest cpu in the nohz.cpu_mask.
4010 int ilb = first_cpu(nohz.cpu_mask);
4012 if (ilb < nr_cpu_ids)
4018 * If this cpu is idle and doing idle load balancing for all the
4019 * cpus with ticks stopped, is it time for that to stop?
4021 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4022 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4028 * If this cpu is idle and the idle load balancing is done by
4029 * someone else, then no need raise the SCHED_SOFTIRQ
4031 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4032 cpu_isset(cpu, nohz.cpu_mask))
4035 if (time_after_eq(jiffies, rq->next_balance))
4036 raise_softirq(SCHED_SOFTIRQ);
4039 #else /* CONFIG_SMP */
4042 * on UP we do not need to balance between CPUs:
4044 static inline void idle_balance(int cpu, struct rq *rq)
4050 DEFINE_PER_CPU(struct kernel_stat, kstat);
4052 EXPORT_PER_CPU_SYMBOL(kstat);
4055 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4056 * that have not yet been banked in case the task is currently running.
4058 unsigned long long task_sched_runtime(struct task_struct *p)
4060 unsigned long flags;
4064 rq = task_rq_lock(p, &flags);
4065 ns = p->se.sum_exec_runtime;
4066 if (task_current(rq, p)) {
4067 update_rq_clock(rq);
4068 delta_exec = rq->clock - p->se.exec_start;
4069 if ((s64)delta_exec > 0)
4072 task_rq_unlock(rq, &flags);
4078 * Account user cpu time to a process.
4079 * @p: the process that the cpu time gets accounted to
4080 * @cputime: the cpu time spent in user space since the last update
4082 void account_user_time(struct task_struct *p, cputime_t cputime)
4084 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4087 p->utime = cputime_add(p->utime, cputime);
4089 /* Add user time to cpustat. */
4090 tmp = cputime_to_cputime64(cputime);
4091 if (TASK_NICE(p) > 0)
4092 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4094 cpustat->user = cputime64_add(cpustat->user, tmp);
4095 /* Account for user time used */
4096 acct_update_integrals(p);
4100 * Account guest cpu time to a process.
4101 * @p: the process that the cpu time gets accounted to
4102 * @cputime: the cpu time spent in virtual machine since the last update
4104 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4107 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4109 tmp = cputime_to_cputime64(cputime);
4111 p->utime = cputime_add(p->utime, cputime);
4112 p->gtime = cputime_add(p->gtime, cputime);
4114 cpustat->user = cputime64_add(cpustat->user, tmp);
4115 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4119 * Account scaled user cpu time to a process.
4120 * @p: the process that the cpu time gets accounted to
4121 * @cputime: the cpu time spent in user space since the last update
4123 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4125 p->utimescaled = cputime_add(p->utimescaled, cputime);
4129 * Account system cpu time to a process.
4130 * @p: the process that the cpu time gets accounted to
4131 * @hardirq_offset: the offset to subtract from hardirq_count()
4132 * @cputime: the cpu time spent in kernel space since the last update
4134 void account_system_time(struct task_struct *p, int hardirq_offset,
4137 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4138 struct rq *rq = this_rq();
4141 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4142 account_guest_time(p, cputime);
4146 p->stime = cputime_add(p->stime, cputime);
4148 /* Add system time to cpustat. */
4149 tmp = cputime_to_cputime64(cputime);
4150 if (hardirq_count() - hardirq_offset)
4151 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4152 else if (softirq_count())
4153 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4154 else if (p != rq->idle)
4155 cpustat->system = cputime64_add(cpustat->system, tmp);
4156 else if (atomic_read(&rq->nr_iowait) > 0)
4157 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4159 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4160 /* Account for system time used */
4161 acct_update_integrals(p);
4165 * Account scaled system cpu time to a process.
4166 * @p: the process that the cpu time gets accounted to
4167 * @hardirq_offset: the offset to subtract from hardirq_count()
4168 * @cputime: the cpu time spent in kernel space since the last update
4170 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4172 p->stimescaled = cputime_add(p->stimescaled, cputime);
4176 * Account for involuntary wait time.
4177 * @p: the process from which the cpu time has been stolen
4178 * @steal: the cpu time spent in involuntary wait
4180 void account_steal_time(struct task_struct *p, cputime_t steal)
4182 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4183 cputime64_t tmp = cputime_to_cputime64(steal);
4184 struct rq *rq = this_rq();
4186 if (p == rq->idle) {
4187 p->stime = cputime_add(p->stime, steal);
4188 if (atomic_read(&rq->nr_iowait) > 0)
4189 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4191 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4193 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4197 * Use precise platform statistics if available:
4199 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4200 cputime_t task_utime(struct task_struct *p)
4205 cputime_t task_stime(struct task_struct *p)
4210 cputime_t task_utime(struct task_struct *p)
4212 clock_t utime = cputime_to_clock_t(p->utime),
4213 total = utime + cputime_to_clock_t(p->stime);
4217 * Use CFS's precise accounting:
4219 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4223 do_div(temp, total);
4225 utime = (clock_t)temp;
4227 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4228 return p->prev_utime;
4231 cputime_t task_stime(struct task_struct *p)
4236 * Use CFS's precise accounting. (we subtract utime from
4237 * the total, to make sure the total observed by userspace
4238 * grows monotonically - apps rely on that):
4240 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4241 cputime_to_clock_t(task_utime(p));
4244 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4246 return p->prev_stime;
4250 inline cputime_t task_gtime(struct task_struct *p)
4256 * This function gets called by the timer code, with HZ frequency.
4257 * We call it with interrupts disabled.
4259 * It also gets called by the fork code, when changing the parent's
4262 void scheduler_tick(void)
4264 int cpu = smp_processor_id();
4265 struct rq *rq = cpu_rq(cpu);
4266 struct task_struct *curr = rq->curr;
4270 spin_lock(&rq->lock);
4271 update_rq_clock(rq);
4272 update_cpu_load(rq);
4273 curr->sched_class->task_tick(rq, curr, 0);
4274 spin_unlock(&rq->lock);
4277 rq->idle_at_tick = idle_cpu(cpu);
4278 trigger_load_balance(rq, cpu);
4282 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4283 defined(CONFIG_PREEMPT_TRACER))
4285 static inline unsigned long get_parent_ip(unsigned long addr)
4287 if (in_lock_functions(addr)) {
4288 addr = CALLER_ADDR2;
4289 if (in_lock_functions(addr))
4290 addr = CALLER_ADDR3;
4295 void __kprobes add_preempt_count(int val)
4297 #ifdef CONFIG_DEBUG_PREEMPT
4301 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4304 preempt_count() += val;
4305 #ifdef CONFIG_DEBUG_PREEMPT
4307 * Spinlock count overflowing soon?
4309 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4312 if (preempt_count() == val)
4313 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4315 EXPORT_SYMBOL(add_preempt_count);
4317 void __kprobes sub_preempt_count(int val)
4319 #ifdef CONFIG_DEBUG_PREEMPT
4323 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4326 * Is the spinlock portion underflowing?
4328 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4329 !(preempt_count() & PREEMPT_MASK)))
4333 if (preempt_count() == val)
4334 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4335 preempt_count() -= val;
4337 EXPORT_SYMBOL(sub_preempt_count);
4342 * Print scheduling while atomic bug:
4344 static noinline void __schedule_bug(struct task_struct *prev)
4346 struct pt_regs *regs = get_irq_regs();
4348 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4349 prev->comm, prev->pid, preempt_count());
4351 debug_show_held_locks(prev);
4353 if (irqs_disabled())
4354 print_irqtrace_events(prev);
4363 * Various schedule()-time debugging checks and statistics:
4365 static inline void schedule_debug(struct task_struct *prev)
4368 * Test if we are atomic. Since do_exit() needs to call into
4369 * schedule() atomically, we ignore that path for now.
4370 * Otherwise, whine if we are scheduling when we should not be.
4372 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4373 __schedule_bug(prev);
4375 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4377 schedstat_inc(this_rq(), sched_count);
4378 #ifdef CONFIG_SCHEDSTATS
4379 if (unlikely(prev->lock_depth >= 0)) {
4380 schedstat_inc(this_rq(), bkl_count);
4381 schedstat_inc(prev, sched_info.bkl_count);
4387 * Pick up the highest-prio task:
4389 static inline struct task_struct *
4390 pick_next_task(struct rq *rq, struct task_struct *prev)
4392 const struct sched_class *class;
4393 struct task_struct *p;
4396 * Optimization: we know that if all tasks are in
4397 * the fair class we can call that function directly:
4399 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4400 p = fair_sched_class.pick_next_task(rq);
4405 class = sched_class_highest;
4407 p = class->pick_next_task(rq);
4411 * Will never be NULL as the idle class always
4412 * returns a non-NULL p:
4414 class = class->next;
4419 * schedule() is the main scheduler function.
4421 asmlinkage void __sched schedule(void)
4423 struct task_struct *prev, *next;
4424 unsigned long *switch_count;
4430 cpu = smp_processor_id();
4434 switch_count = &prev->nivcsw;
4436 release_kernel_lock(prev);
4437 need_resched_nonpreemptible:
4439 schedule_debug(prev);
4441 if (sched_feat(HRTICK))
4444 spin_lock_irq(&rq->lock);
4445 update_rq_clock(rq);
4446 clear_tsk_need_resched(prev);
4448 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4449 if (unlikely(signal_pending_state(prev->state, prev)))
4450 prev->state = TASK_RUNNING;
4452 deactivate_task(rq, prev, 1);
4453 switch_count = &prev->nvcsw;
4457 if (prev->sched_class->pre_schedule)
4458 prev->sched_class->pre_schedule(rq, prev);
4461 if (unlikely(!rq->nr_running))
4462 idle_balance(cpu, rq);
4464 prev->sched_class->put_prev_task(rq, prev);
4465 next = pick_next_task(rq, prev);
4467 if (likely(prev != next)) {
4468 sched_info_switch(prev, next);
4474 context_switch(rq, prev, next); /* unlocks the rq */
4476 * the context switch might have flipped the stack from under
4477 * us, hence refresh the local variables.
4479 cpu = smp_processor_id();
4482 spin_unlock_irq(&rq->lock);
4484 if (unlikely(reacquire_kernel_lock(current) < 0))
4485 goto need_resched_nonpreemptible;
4487 preempt_enable_no_resched();
4488 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4491 EXPORT_SYMBOL(schedule);
4493 #ifdef CONFIG_PREEMPT
4495 * this is the entry point to schedule() from in-kernel preemption
4496 * off of preempt_enable. Kernel preemptions off return from interrupt
4497 * occur there and call schedule directly.
4499 asmlinkage void __sched preempt_schedule(void)
4501 struct thread_info *ti = current_thread_info();
4504 * If there is a non-zero preempt_count or interrupts are disabled,
4505 * we do not want to preempt the current task. Just return..
4507 if (likely(ti->preempt_count || irqs_disabled()))
4511 add_preempt_count(PREEMPT_ACTIVE);
4513 sub_preempt_count(PREEMPT_ACTIVE);
4516 * Check again in case we missed a preemption opportunity
4517 * between schedule and now.
4520 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4522 EXPORT_SYMBOL(preempt_schedule);
4525 * this is the entry point to schedule() from kernel preemption
4526 * off of irq context.
4527 * Note, that this is called and return with irqs disabled. This will
4528 * protect us against recursive calling from irq.
4530 asmlinkage void __sched preempt_schedule_irq(void)
4532 struct thread_info *ti = current_thread_info();
4534 /* Catch callers which need to be fixed */
4535 BUG_ON(ti->preempt_count || !irqs_disabled());
4538 add_preempt_count(PREEMPT_ACTIVE);
4541 local_irq_disable();
4542 sub_preempt_count(PREEMPT_ACTIVE);
4545 * Check again in case we missed a preemption opportunity
4546 * between schedule and now.
4549 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4552 #endif /* CONFIG_PREEMPT */
4554 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4557 return try_to_wake_up(curr->private, mode, sync);
4559 EXPORT_SYMBOL(default_wake_function);
4562 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4563 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4564 * number) then we wake all the non-exclusive tasks and one exclusive task.
4566 * There are circumstances in which we can try to wake a task which has already
4567 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4568 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4570 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4571 int nr_exclusive, int sync, void *key)
4573 wait_queue_t *curr, *next;
4575 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4576 unsigned flags = curr->flags;
4578 if (curr->func(curr, mode, sync, key) &&
4579 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4585 * __wake_up - wake up threads blocked on a waitqueue.
4587 * @mode: which threads
4588 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4589 * @key: is directly passed to the wakeup function
4591 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4592 int nr_exclusive, void *key)
4594 unsigned long flags;
4596 spin_lock_irqsave(&q->lock, flags);
4597 __wake_up_common(q, mode, nr_exclusive, 0, key);
4598 spin_unlock_irqrestore(&q->lock, flags);
4600 EXPORT_SYMBOL(__wake_up);
4603 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4605 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4607 __wake_up_common(q, mode, 1, 0, NULL);
4611 * __wake_up_sync - wake up threads blocked on a waitqueue.
4613 * @mode: which threads
4614 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4616 * The sync wakeup differs that the waker knows that it will schedule
4617 * away soon, so while the target thread will be woken up, it will not
4618 * be migrated to another CPU - ie. the two threads are 'synchronized'
4619 * with each other. This can prevent needless bouncing between CPUs.
4621 * On UP it can prevent extra preemption.
4624 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4626 unsigned long flags;
4632 if (unlikely(!nr_exclusive))
4635 spin_lock_irqsave(&q->lock, flags);
4636 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4637 spin_unlock_irqrestore(&q->lock, flags);
4639 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4642 * complete: - signals a single thread waiting on this completion
4643 * @x: holds the state of this particular completion
4645 * This will wake up a single thread waiting on this completion. Threads will be
4646 * awakened in the same order in which they were queued.
4648 * See also complete_all(), wait_for_completion() and related routines.
4650 void complete(struct completion *x)
4652 unsigned long flags;
4654 spin_lock_irqsave(&x->wait.lock, flags);
4656 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4657 spin_unlock_irqrestore(&x->wait.lock, flags);
4659 EXPORT_SYMBOL(complete);
4662 * complete_all: - signals all threads waiting on this completion
4663 * @x: holds the state of this particular completion
4665 * This will wake up all threads waiting on this particular completion event.
4667 void complete_all(struct completion *x)
4669 unsigned long flags;
4671 spin_lock_irqsave(&x->wait.lock, flags);
4672 x->done += UINT_MAX/2;
4673 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4674 spin_unlock_irqrestore(&x->wait.lock, flags);
4676 EXPORT_SYMBOL(complete_all);
4678 static inline long __sched
4679 do_wait_for_common(struct completion *x, long timeout, int state)
4682 DECLARE_WAITQUEUE(wait, current);
4684 wait.flags |= WQ_FLAG_EXCLUSIVE;
4685 __add_wait_queue_tail(&x->wait, &wait);
4687 if (signal_pending_state(state, current)) {
4688 timeout = -ERESTARTSYS;
4691 __set_current_state(state);
4692 spin_unlock_irq(&x->wait.lock);
4693 timeout = schedule_timeout(timeout);
4694 spin_lock_irq(&x->wait.lock);
4695 } while (!x->done && timeout);
4696 __remove_wait_queue(&x->wait, &wait);
4701 return timeout ?: 1;
4705 wait_for_common(struct completion *x, long timeout, int state)
4709 spin_lock_irq(&x->wait.lock);
4710 timeout = do_wait_for_common(x, timeout, state);
4711 spin_unlock_irq(&x->wait.lock);
4716 * wait_for_completion: - waits for completion of a task
4717 * @x: holds the state of this particular completion
4719 * This waits to be signaled for completion of a specific task. It is NOT
4720 * interruptible and there is no timeout.
4722 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4723 * and interrupt capability. Also see complete().
4725 void __sched wait_for_completion(struct completion *x)
4727 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4729 EXPORT_SYMBOL(wait_for_completion);
4732 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4733 * @x: holds the state of this particular completion
4734 * @timeout: timeout value in jiffies
4736 * This waits for either a completion of a specific task to be signaled or for a
4737 * specified timeout to expire. The timeout is in jiffies. It is not
4740 unsigned long __sched
4741 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4743 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4745 EXPORT_SYMBOL(wait_for_completion_timeout);
4748 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4749 * @x: holds the state of this particular completion
4751 * This waits for completion of a specific task to be signaled. It is
4754 int __sched wait_for_completion_interruptible(struct completion *x)
4756 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4757 if (t == -ERESTARTSYS)
4761 EXPORT_SYMBOL(wait_for_completion_interruptible);
4764 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4765 * @x: holds the state of this particular completion
4766 * @timeout: timeout value in jiffies
4768 * This waits for either a completion of a specific task to be signaled or for a
4769 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4771 unsigned long __sched
4772 wait_for_completion_interruptible_timeout(struct completion *x,
4773 unsigned long timeout)
4775 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4777 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4780 * wait_for_completion_killable: - waits for completion of a task (killable)
4781 * @x: holds the state of this particular completion
4783 * This waits to be signaled for completion of a specific task. It can be
4784 * interrupted by a kill signal.
4786 int __sched wait_for_completion_killable(struct completion *x)
4788 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4789 if (t == -ERESTARTSYS)
4793 EXPORT_SYMBOL(wait_for_completion_killable);
4796 * try_wait_for_completion - try to decrement a completion without blocking
4797 * @x: completion structure
4799 * Returns: 0 if a decrement cannot be done without blocking
4800 * 1 if a decrement succeeded.
4802 * If a completion is being used as a counting completion,
4803 * attempt to decrement the counter without blocking. This
4804 * enables us to avoid waiting if the resource the completion
4805 * is protecting is not available.
4807 bool try_wait_for_completion(struct completion *x)
4811 spin_lock_irq(&x->wait.lock);
4816 spin_unlock_irq(&x->wait.lock);
4819 EXPORT_SYMBOL(try_wait_for_completion);
4822 * completion_done - Test to see if a completion has any waiters
4823 * @x: completion structure
4825 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4826 * 1 if there are no waiters.
4829 bool completion_done(struct completion *x)
4833 spin_lock_irq(&x->wait.lock);
4836 spin_unlock_irq(&x->wait.lock);
4839 EXPORT_SYMBOL(completion_done);
4842 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4844 unsigned long flags;
4847 init_waitqueue_entry(&wait, current);
4849 __set_current_state(state);
4851 spin_lock_irqsave(&q->lock, flags);
4852 __add_wait_queue(q, &wait);
4853 spin_unlock(&q->lock);
4854 timeout = schedule_timeout(timeout);
4855 spin_lock_irq(&q->lock);
4856 __remove_wait_queue(q, &wait);
4857 spin_unlock_irqrestore(&q->lock, flags);
4862 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4864 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4866 EXPORT_SYMBOL(interruptible_sleep_on);
4869 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4871 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4873 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4875 void __sched sleep_on(wait_queue_head_t *q)
4877 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4879 EXPORT_SYMBOL(sleep_on);
4881 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4883 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4885 EXPORT_SYMBOL(sleep_on_timeout);
4887 #ifdef CONFIG_RT_MUTEXES
4890 * rt_mutex_setprio - set the current priority of a task
4892 * @prio: prio value (kernel-internal form)
4894 * This function changes the 'effective' priority of a task. It does
4895 * not touch ->normal_prio like __setscheduler().
4897 * Used by the rt_mutex code to implement priority inheritance logic.
4899 void rt_mutex_setprio(struct task_struct *p, int prio)
4901 unsigned long flags;
4902 int oldprio, on_rq, running;
4904 const struct sched_class *prev_class = p->sched_class;
4906 BUG_ON(prio < 0 || prio > MAX_PRIO);
4908 rq = task_rq_lock(p, &flags);
4909 update_rq_clock(rq);
4912 on_rq = p->se.on_rq;
4913 running = task_current(rq, p);
4915 dequeue_task(rq, p, 0);
4917 p->sched_class->put_prev_task(rq, p);
4920 p->sched_class = &rt_sched_class;
4922 p->sched_class = &fair_sched_class;
4927 p->sched_class->set_curr_task(rq);
4929 enqueue_task(rq, p, 0);
4931 check_class_changed(rq, p, prev_class, oldprio, running);
4933 task_rq_unlock(rq, &flags);
4938 void set_user_nice(struct task_struct *p, long nice)
4940 int old_prio, delta, on_rq;
4941 unsigned long flags;
4944 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4947 * We have to be careful, if called from sys_setpriority(),
4948 * the task might be in the middle of scheduling on another CPU.
4950 rq = task_rq_lock(p, &flags);
4951 update_rq_clock(rq);
4953 * The RT priorities are set via sched_setscheduler(), but we still
4954 * allow the 'normal' nice value to be set - but as expected
4955 * it wont have any effect on scheduling until the task is
4956 * SCHED_FIFO/SCHED_RR:
4958 if (task_has_rt_policy(p)) {
4959 p->static_prio = NICE_TO_PRIO(nice);
4962 on_rq = p->se.on_rq;
4964 dequeue_task(rq, p, 0);
4966 p->static_prio = NICE_TO_PRIO(nice);
4969 p->prio = effective_prio(p);
4970 delta = p->prio - old_prio;
4973 enqueue_task(rq, p, 0);
4975 * If the task increased its priority or is running and
4976 * lowered its priority, then reschedule its CPU:
4978 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4979 resched_task(rq->curr);
4982 task_rq_unlock(rq, &flags);
4984 EXPORT_SYMBOL(set_user_nice);
4987 * can_nice - check if a task can reduce its nice value
4991 int can_nice(const struct task_struct *p, const int nice)
4993 /* convert nice value [19,-20] to rlimit style value [1,40] */
4994 int nice_rlim = 20 - nice;
4996 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4997 capable(CAP_SYS_NICE));
5000 #ifdef __ARCH_WANT_SYS_NICE
5003 * sys_nice - change the priority of the current process.
5004 * @increment: priority increment
5006 * sys_setpriority is a more generic, but much slower function that
5007 * does similar things.
5009 asmlinkage long sys_nice(int increment)
5014 * Setpriority might change our priority at the same moment.
5015 * We don't have to worry. Conceptually one call occurs first
5016 * and we have a single winner.
5018 if (increment < -40)
5023 nice = PRIO_TO_NICE(current->static_prio) + increment;
5029 if (increment < 0 && !can_nice(current, nice))
5032 retval = security_task_setnice(current, nice);
5036 set_user_nice(current, nice);
5043 * task_prio - return the priority value of a given task.
5044 * @p: the task in question.
5046 * This is the priority value as seen by users in /proc.
5047 * RT tasks are offset by -200. Normal tasks are centered
5048 * around 0, value goes from -16 to +15.
5050 int task_prio(const struct task_struct *p)
5052 return p->prio - MAX_RT_PRIO;
5056 * task_nice - return the nice value of a given task.
5057 * @p: the task in question.
5059 int task_nice(const struct task_struct *p)
5061 return TASK_NICE(p);
5063 EXPORT_SYMBOL(task_nice);
5066 * idle_cpu - is a given cpu idle currently?
5067 * @cpu: the processor in question.
5069 int idle_cpu(int cpu)
5071 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5075 * idle_task - return the idle task for a given cpu.
5076 * @cpu: the processor in question.
5078 struct task_struct *idle_task(int cpu)
5080 return cpu_rq(cpu)->idle;
5084 * find_process_by_pid - find a process with a matching PID value.
5085 * @pid: the pid in question.
5087 static struct task_struct *find_process_by_pid(pid_t pid)
5089 return pid ? find_task_by_vpid(pid) : current;
5092 /* Actually do priority change: must hold rq lock. */
5094 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5096 BUG_ON(p->se.on_rq);
5099 switch (p->policy) {
5103 p->sched_class = &fair_sched_class;
5107 p->sched_class = &rt_sched_class;
5111 p->rt_priority = prio;
5112 p->normal_prio = normal_prio(p);
5113 /* we are holding p->pi_lock already */
5114 p->prio = rt_mutex_getprio(p);
5118 static int __sched_setscheduler(struct task_struct *p, int policy,
5119 struct sched_param *param, bool user)
5121 int retval, oldprio, oldpolicy = -1, on_rq, running;
5122 unsigned long flags;
5123 const struct sched_class *prev_class = p->sched_class;
5126 /* may grab non-irq protected spin_locks */
5127 BUG_ON(in_interrupt());
5129 /* double check policy once rq lock held */
5131 policy = oldpolicy = p->policy;
5132 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5133 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5134 policy != SCHED_IDLE)
5137 * Valid priorities for SCHED_FIFO and SCHED_RR are
5138 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5139 * SCHED_BATCH and SCHED_IDLE is 0.
5141 if (param->sched_priority < 0 ||
5142 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5143 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5145 if (rt_policy(policy) != (param->sched_priority != 0))
5149 * Allow unprivileged RT tasks to decrease priority:
5151 if (user && !capable(CAP_SYS_NICE)) {
5152 if (rt_policy(policy)) {
5153 unsigned long rlim_rtprio;
5155 if (!lock_task_sighand(p, &flags))
5157 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5158 unlock_task_sighand(p, &flags);
5160 /* can't set/change the rt policy */
5161 if (policy != p->policy && !rlim_rtprio)
5164 /* can't increase priority */
5165 if (param->sched_priority > p->rt_priority &&
5166 param->sched_priority > rlim_rtprio)
5170 * Like positive nice levels, dont allow tasks to
5171 * move out of SCHED_IDLE either:
5173 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5176 /* can't change other user's priorities */
5177 if ((current->euid != p->euid) &&
5178 (current->euid != p->uid))
5183 #ifdef CONFIG_RT_GROUP_SCHED
5185 * Do not allow realtime tasks into groups that have no runtime
5188 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5189 task_group(p)->rt_bandwidth.rt_runtime == 0)
5193 retval = security_task_setscheduler(p, policy, param);
5199 * make sure no PI-waiters arrive (or leave) while we are
5200 * changing the priority of the task:
5202 spin_lock_irqsave(&p->pi_lock, flags);
5204 * To be able to change p->policy safely, the apropriate
5205 * runqueue lock must be held.
5207 rq = __task_rq_lock(p);
5208 /* recheck policy now with rq lock held */
5209 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5210 policy = oldpolicy = -1;
5211 __task_rq_unlock(rq);
5212 spin_unlock_irqrestore(&p->pi_lock, flags);
5215 update_rq_clock(rq);
5216 on_rq = p->se.on_rq;
5217 running = task_current(rq, p);
5219 deactivate_task(rq, p, 0);
5221 p->sched_class->put_prev_task(rq, p);
5224 __setscheduler(rq, p, policy, param->sched_priority);
5227 p->sched_class->set_curr_task(rq);
5229 activate_task(rq, p, 0);
5231 check_class_changed(rq, p, prev_class, oldprio, running);
5233 __task_rq_unlock(rq);
5234 spin_unlock_irqrestore(&p->pi_lock, flags);
5236 rt_mutex_adjust_pi(p);
5242 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5243 * @p: the task in question.
5244 * @policy: new policy.
5245 * @param: structure containing the new RT priority.
5247 * NOTE that the task may be already dead.
5249 int sched_setscheduler(struct task_struct *p, int policy,
5250 struct sched_param *param)
5252 return __sched_setscheduler(p, policy, param, true);
5254 EXPORT_SYMBOL_GPL(sched_setscheduler);
5257 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5258 * @p: the task in question.
5259 * @policy: new policy.
5260 * @param: structure containing the new RT priority.
5262 * Just like sched_setscheduler, only don't bother checking if the
5263 * current context has permission. For example, this is needed in
5264 * stop_machine(): we create temporary high priority worker threads,
5265 * but our caller might not have that capability.
5267 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5268 struct sched_param *param)
5270 return __sched_setscheduler(p, policy, param, false);
5274 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5276 struct sched_param lparam;
5277 struct task_struct *p;
5280 if (!param || pid < 0)
5282 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5287 p = find_process_by_pid(pid);
5289 retval = sched_setscheduler(p, policy, &lparam);
5296 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5297 * @pid: the pid in question.
5298 * @policy: new policy.
5299 * @param: structure containing the new RT priority.
5302 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5304 /* negative values for policy are not valid */
5308 return do_sched_setscheduler(pid, policy, param);
5312 * sys_sched_setparam - set/change the RT priority of a thread
5313 * @pid: the pid in question.
5314 * @param: structure containing the new RT priority.
5316 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5318 return do_sched_setscheduler(pid, -1, param);
5322 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5323 * @pid: the pid in question.
5325 asmlinkage long sys_sched_getscheduler(pid_t pid)
5327 struct task_struct *p;
5334 read_lock(&tasklist_lock);
5335 p = find_process_by_pid(pid);
5337 retval = security_task_getscheduler(p);
5341 read_unlock(&tasklist_lock);
5346 * sys_sched_getscheduler - get the RT priority of a thread
5347 * @pid: the pid in question.
5348 * @param: structure containing the RT priority.
5350 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5352 struct sched_param lp;
5353 struct task_struct *p;
5356 if (!param || pid < 0)
5359 read_lock(&tasklist_lock);
5360 p = find_process_by_pid(pid);
5365 retval = security_task_getscheduler(p);
5369 lp.sched_priority = p->rt_priority;
5370 read_unlock(&tasklist_lock);
5373 * This one might sleep, we cannot do it with a spinlock held ...
5375 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5380 read_unlock(&tasklist_lock);
5384 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5386 cpumask_t cpus_allowed;
5387 cpumask_t new_mask = *in_mask;
5388 struct task_struct *p;
5392 read_lock(&tasklist_lock);
5394 p = find_process_by_pid(pid);
5396 read_unlock(&tasklist_lock);
5402 * It is not safe to call set_cpus_allowed with the
5403 * tasklist_lock held. We will bump the task_struct's
5404 * usage count and then drop tasklist_lock.
5407 read_unlock(&tasklist_lock);
5410 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5411 !capable(CAP_SYS_NICE))
5414 retval = security_task_setscheduler(p, 0, NULL);
5418 cpuset_cpus_allowed(p, &cpus_allowed);
5419 cpus_and(new_mask, new_mask, cpus_allowed);
5421 retval = set_cpus_allowed_ptr(p, &new_mask);
5424 cpuset_cpus_allowed(p, &cpus_allowed);
5425 if (!cpus_subset(new_mask, cpus_allowed)) {
5427 * We must have raced with a concurrent cpuset
5428 * update. Just reset the cpus_allowed to the
5429 * cpuset's cpus_allowed
5431 new_mask = cpus_allowed;
5441 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5442 cpumask_t *new_mask)
5444 if (len < sizeof(cpumask_t)) {
5445 memset(new_mask, 0, sizeof(cpumask_t));
5446 } else if (len > sizeof(cpumask_t)) {
5447 len = sizeof(cpumask_t);
5449 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5453 * sys_sched_setaffinity - set the cpu affinity of a process
5454 * @pid: pid of the process
5455 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5456 * @user_mask_ptr: user-space pointer to the new cpu mask
5458 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5459 unsigned long __user *user_mask_ptr)
5464 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5468 return sched_setaffinity(pid, &new_mask);
5471 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5473 struct task_struct *p;
5477 read_lock(&tasklist_lock);
5480 p = find_process_by_pid(pid);
5484 retval = security_task_getscheduler(p);
5488 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5491 read_unlock(&tasklist_lock);
5498 * sys_sched_getaffinity - get the cpu affinity of a process
5499 * @pid: pid of the process
5500 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5501 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5503 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5504 unsigned long __user *user_mask_ptr)
5509 if (len < sizeof(cpumask_t))
5512 ret = sched_getaffinity(pid, &mask);
5516 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5519 return sizeof(cpumask_t);
5523 * sys_sched_yield - yield the current processor to other threads.
5525 * This function yields the current CPU to other tasks. If there are no
5526 * other threads running on this CPU then this function will return.
5528 asmlinkage long sys_sched_yield(void)
5530 struct rq *rq = this_rq_lock();
5532 schedstat_inc(rq, yld_count);
5533 current->sched_class->yield_task(rq);
5536 * Since we are going to call schedule() anyway, there's
5537 * no need to preempt or enable interrupts:
5539 __release(rq->lock);
5540 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5541 _raw_spin_unlock(&rq->lock);
5542 preempt_enable_no_resched();
5549 static void __cond_resched(void)
5551 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5552 __might_sleep(__FILE__, __LINE__);
5555 * The BKS might be reacquired before we have dropped
5556 * PREEMPT_ACTIVE, which could trigger a second
5557 * cond_resched() call.
5560 add_preempt_count(PREEMPT_ACTIVE);
5562 sub_preempt_count(PREEMPT_ACTIVE);
5563 } while (need_resched());
5566 int __sched _cond_resched(void)
5568 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5569 system_state == SYSTEM_RUNNING) {
5575 EXPORT_SYMBOL(_cond_resched);
5578 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5579 * call schedule, and on return reacquire the lock.
5581 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5582 * operations here to prevent schedule() from being called twice (once via
5583 * spin_unlock(), once by hand).
5585 int cond_resched_lock(spinlock_t *lock)
5587 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5590 if (spin_needbreak(lock) || resched) {
5592 if (resched && need_resched())
5601 EXPORT_SYMBOL(cond_resched_lock);
5603 int __sched cond_resched_softirq(void)
5605 BUG_ON(!in_softirq());
5607 if (need_resched() && system_state == SYSTEM_RUNNING) {
5615 EXPORT_SYMBOL(cond_resched_softirq);
5618 * yield - yield the current processor to other threads.
5620 * This is a shortcut for kernel-space yielding - it marks the
5621 * thread runnable and calls sys_sched_yield().
5623 void __sched yield(void)
5625 set_current_state(TASK_RUNNING);
5628 EXPORT_SYMBOL(yield);
5631 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5632 * that process accounting knows that this is a task in IO wait state.
5634 * But don't do that if it is a deliberate, throttling IO wait (this task
5635 * has set its backing_dev_info: the queue against which it should throttle)
5637 void __sched io_schedule(void)
5639 struct rq *rq = &__raw_get_cpu_var(runqueues);
5641 delayacct_blkio_start();
5642 atomic_inc(&rq->nr_iowait);
5644 atomic_dec(&rq->nr_iowait);
5645 delayacct_blkio_end();
5647 EXPORT_SYMBOL(io_schedule);
5649 long __sched io_schedule_timeout(long timeout)
5651 struct rq *rq = &__raw_get_cpu_var(runqueues);
5654 delayacct_blkio_start();
5655 atomic_inc(&rq->nr_iowait);
5656 ret = schedule_timeout(timeout);
5657 atomic_dec(&rq->nr_iowait);
5658 delayacct_blkio_end();
5663 * sys_sched_get_priority_max - return maximum RT priority.
5664 * @policy: scheduling class.
5666 * this syscall returns the maximum rt_priority that can be used
5667 * by a given scheduling class.
5669 asmlinkage long sys_sched_get_priority_max(int policy)
5676 ret = MAX_USER_RT_PRIO-1;
5688 * sys_sched_get_priority_min - return minimum RT priority.
5689 * @policy: scheduling class.
5691 * this syscall returns the minimum rt_priority that can be used
5692 * by a given scheduling class.
5694 asmlinkage long sys_sched_get_priority_min(int policy)
5712 * sys_sched_rr_get_interval - return the default timeslice of a process.
5713 * @pid: pid of the process.
5714 * @interval: userspace pointer to the timeslice value.
5716 * this syscall writes the default timeslice value of a given process
5717 * into the user-space timespec buffer. A value of '0' means infinity.
5720 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5722 struct task_struct *p;
5723 unsigned int time_slice;
5731 read_lock(&tasklist_lock);
5732 p = find_process_by_pid(pid);
5736 retval = security_task_getscheduler(p);
5741 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5742 * tasks that are on an otherwise idle runqueue:
5745 if (p->policy == SCHED_RR) {
5746 time_slice = DEF_TIMESLICE;
5747 } else if (p->policy != SCHED_FIFO) {
5748 struct sched_entity *se = &p->se;
5749 unsigned long flags;
5752 rq = task_rq_lock(p, &flags);
5753 if (rq->cfs.load.weight)
5754 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5755 task_rq_unlock(rq, &flags);
5757 read_unlock(&tasklist_lock);
5758 jiffies_to_timespec(time_slice, &t);
5759 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5763 read_unlock(&tasklist_lock);
5767 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5769 void sched_show_task(struct task_struct *p)
5771 unsigned long free = 0;
5774 state = p->state ? __ffs(p->state) + 1 : 0;
5775 printk(KERN_INFO "%-13.13s %c", p->comm,
5776 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5777 #if BITS_PER_LONG == 32
5778 if (state == TASK_RUNNING)
5779 printk(KERN_CONT " running ");
5781 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5783 if (state == TASK_RUNNING)
5784 printk(KERN_CONT " running task ");
5786 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5788 #ifdef CONFIG_DEBUG_STACK_USAGE
5790 unsigned long *n = end_of_stack(p);
5793 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5796 printk(KERN_CONT "%5lu %5d %6d\n", free,
5797 task_pid_nr(p), task_pid_nr(p->real_parent));
5799 show_stack(p, NULL);
5802 void show_state_filter(unsigned long state_filter)
5804 struct task_struct *g, *p;
5806 #if BITS_PER_LONG == 32
5808 " task PC stack pid father\n");
5811 " task PC stack pid father\n");
5813 read_lock(&tasklist_lock);
5814 do_each_thread(g, p) {
5816 * reset the NMI-timeout, listing all files on a slow
5817 * console might take alot of time:
5819 touch_nmi_watchdog();
5820 if (!state_filter || (p->state & state_filter))
5822 } while_each_thread(g, p);
5824 touch_all_softlockup_watchdogs();
5826 #ifdef CONFIG_SCHED_DEBUG
5827 sysrq_sched_debug_show();
5829 read_unlock(&tasklist_lock);
5831 * Only show locks if all tasks are dumped:
5833 if (state_filter == -1)
5834 debug_show_all_locks();
5837 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5839 idle->sched_class = &idle_sched_class;
5843 * init_idle - set up an idle thread for a given CPU
5844 * @idle: task in question
5845 * @cpu: cpu the idle task belongs to
5847 * NOTE: this function does not set the idle thread's NEED_RESCHED
5848 * flag, to make booting more robust.
5850 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5852 struct rq *rq = cpu_rq(cpu);
5853 unsigned long flags;
5856 idle->se.exec_start = sched_clock();
5858 idle->prio = idle->normal_prio = MAX_PRIO;
5859 idle->cpus_allowed = cpumask_of_cpu(cpu);
5860 __set_task_cpu(idle, cpu);
5862 spin_lock_irqsave(&rq->lock, flags);
5863 rq->curr = rq->idle = idle;
5864 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5867 spin_unlock_irqrestore(&rq->lock, flags);
5869 /* Set the preempt count _outside_ the spinlocks! */
5870 #if defined(CONFIG_PREEMPT)
5871 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5873 task_thread_info(idle)->preempt_count = 0;
5876 * The idle tasks have their own, simple scheduling class:
5878 idle->sched_class = &idle_sched_class;
5882 * In a system that switches off the HZ timer nohz_cpu_mask
5883 * indicates which cpus entered this state. This is used
5884 * in the rcu update to wait only for active cpus. For system
5885 * which do not switch off the HZ timer nohz_cpu_mask should
5886 * always be CPU_MASK_NONE.
5888 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5891 * Increase the granularity value when there are more CPUs,
5892 * because with more CPUs the 'effective latency' as visible
5893 * to users decreases. But the relationship is not linear,
5894 * so pick a second-best guess by going with the log2 of the
5897 * This idea comes from the SD scheduler of Con Kolivas:
5899 static inline void sched_init_granularity(void)
5901 unsigned int factor = 1 + ilog2(num_online_cpus());
5902 const unsigned long limit = 200000000;
5904 sysctl_sched_min_granularity *= factor;
5905 if (sysctl_sched_min_granularity > limit)
5906 sysctl_sched_min_granularity = limit;
5908 sysctl_sched_latency *= factor;
5909 if (sysctl_sched_latency > limit)
5910 sysctl_sched_latency = limit;
5912 sysctl_sched_wakeup_granularity *= factor;
5914 sysctl_sched_shares_ratelimit *= factor;
5919 * This is how migration works:
5921 * 1) we queue a struct migration_req structure in the source CPU's
5922 * runqueue and wake up that CPU's migration thread.
5923 * 2) we down() the locked semaphore => thread blocks.
5924 * 3) migration thread wakes up (implicitly it forces the migrated
5925 * thread off the CPU)
5926 * 4) it gets the migration request and checks whether the migrated
5927 * task is still in the wrong runqueue.
5928 * 5) if it's in the wrong runqueue then the migration thread removes
5929 * it and puts it into the right queue.
5930 * 6) migration thread up()s the semaphore.
5931 * 7) we wake up and the migration is done.
5935 * Change a given task's CPU affinity. Migrate the thread to a
5936 * proper CPU and schedule it away if the CPU it's executing on
5937 * is removed from the allowed bitmask.
5939 * NOTE: the caller must have a valid reference to the task, the
5940 * task must not exit() & deallocate itself prematurely. The
5941 * call is not atomic; no spinlocks may be held.
5943 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5945 struct migration_req req;
5946 unsigned long flags;
5950 rq = task_rq_lock(p, &flags);
5951 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5956 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5957 !cpus_equal(p->cpus_allowed, *new_mask))) {
5962 if (p->sched_class->set_cpus_allowed)
5963 p->sched_class->set_cpus_allowed(p, new_mask);
5965 p->cpus_allowed = *new_mask;
5966 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5969 /* Can the task run on the task's current CPU? If so, we're done */
5970 if (cpu_isset(task_cpu(p), *new_mask))
5973 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5974 /* Need help from migration thread: drop lock and wait. */
5975 task_rq_unlock(rq, &flags);
5976 wake_up_process(rq->migration_thread);
5977 wait_for_completion(&req.done);
5978 tlb_migrate_finish(p->mm);
5982 task_rq_unlock(rq, &flags);
5986 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5989 * Move (not current) task off this cpu, onto dest cpu. We're doing
5990 * this because either it can't run here any more (set_cpus_allowed()
5991 * away from this CPU, or CPU going down), or because we're
5992 * attempting to rebalance this task on exec (sched_exec).
5994 * So we race with normal scheduler movements, but that's OK, as long
5995 * as the task is no longer on this CPU.
5997 * Returns non-zero if task was successfully migrated.
5999 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6001 struct rq *rq_dest, *rq_src;
6004 if (unlikely(!cpu_active(dest_cpu)))
6007 rq_src = cpu_rq(src_cpu);
6008 rq_dest = cpu_rq(dest_cpu);
6010 double_rq_lock(rq_src, rq_dest);
6011 /* Already moved. */
6012 if (task_cpu(p) != src_cpu)
6014 /* Affinity changed (again). */
6015 if (!cpu_isset(dest_cpu, p->cpus_allowed))
6018 on_rq = p->se.on_rq;
6020 deactivate_task(rq_src, p, 0);
6022 set_task_cpu(p, dest_cpu);
6024 activate_task(rq_dest, p, 0);
6025 check_preempt_curr(rq_dest, p, 0);
6030 double_rq_unlock(rq_src, rq_dest);
6035 * migration_thread - this is a highprio system thread that performs
6036 * thread migration by bumping thread off CPU then 'pushing' onto
6039 static int migration_thread(void *data)
6041 int cpu = (long)data;
6045 BUG_ON(rq->migration_thread != current);
6047 set_current_state(TASK_INTERRUPTIBLE);
6048 while (!kthread_should_stop()) {
6049 struct migration_req *req;
6050 struct list_head *head;
6052 spin_lock_irq(&rq->lock);
6054 if (cpu_is_offline(cpu)) {
6055 spin_unlock_irq(&rq->lock);
6059 if (rq->active_balance) {
6060 active_load_balance(rq, cpu);
6061 rq->active_balance = 0;
6064 head = &rq->migration_queue;
6066 if (list_empty(head)) {
6067 spin_unlock_irq(&rq->lock);
6069 set_current_state(TASK_INTERRUPTIBLE);
6072 req = list_entry(head->next, struct migration_req, list);
6073 list_del_init(head->next);
6075 spin_unlock(&rq->lock);
6076 __migrate_task(req->task, cpu, req->dest_cpu);
6079 complete(&req->done);
6081 __set_current_state(TASK_RUNNING);
6085 /* Wait for kthread_stop */
6086 set_current_state(TASK_INTERRUPTIBLE);
6087 while (!kthread_should_stop()) {
6089 set_current_state(TASK_INTERRUPTIBLE);
6091 __set_current_state(TASK_RUNNING);
6095 #ifdef CONFIG_HOTPLUG_CPU
6097 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6101 local_irq_disable();
6102 ret = __migrate_task(p, src_cpu, dest_cpu);
6108 * Figure out where task on dead CPU should go, use force if necessary.
6109 * NOTE: interrupts should be disabled by the caller
6111 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6113 unsigned long flags;
6120 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6121 cpus_and(mask, mask, p->cpus_allowed);
6122 dest_cpu = any_online_cpu(mask);
6124 /* On any allowed CPU? */
6125 if (dest_cpu >= nr_cpu_ids)
6126 dest_cpu = any_online_cpu(p->cpus_allowed);
6128 /* No more Mr. Nice Guy. */
6129 if (dest_cpu >= nr_cpu_ids) {
6130 cpumask_t cpus_allowed;
6132 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6134 * Try to stay on the same cpuset, where the
6135 * current cpuset may be a subset of all cpus.
6136 * The cpuset_cpus_allowed_locked() variant of
6137 * cpuset_cpus_allowed() will not block. It must be
6138 * called within calls to cpuset_lock/cpuset_unlock.
6140 rq = task_rq_lock(p, &flags);
6141 p->cpus_allowed = cpus_allowed;
6142 dest_cpu = any_online_cpu(p->cpus_allowed);
6143 task_rq_unlock(rq, &flags);
6146 * Don't tell them about moving exiting tasks or
6147 * kernel threads (both mm NULL), since they never
6150 if (p->mm && printk_ratelimit()) {
6151 printk(KERN_INFO "process %d (%s) no "
6152 "longer affine to cpu%d\n",
6153 task_pid_nr(p), p->comm, dead_cpu);
6156 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6160 * While a dead CPU has no uninterruptible tasks queued at this point,
6161 * it might still have a nonzero ->nr_uninterruptible counter, because
6162 * for performance reasons the counter is not stricly tracking tasks to
6163 * their home CPUs. So we just add the counter to another CPU's counter,
6164 * to keep the global sum constant after CPU-down:
6166 static void migrate_nr_uninterruptible(struct rq *rq_src)
6168 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6169 unsigned long flags;
6171 local_irq_save(flags);
6172 double_rq_lock(rq_src, rq_dest);
6173 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6174 rq_src->nr_uninterruptible = 0;
6175 double_rq_unlock(rq_src, rq_dest);
6176 local_irq_restore(flags);
6179 /* Run through task list and migrate tasks from the dead cpu. */
6180 static void migrate_live_tasks(int src_cpu)
6182 struct task_struct *p, *t;
6184 read_lock(&tasklist_lock);
6186 do_each_thread(t, p) {
6190 if (task_cpu(p) == src_cpu)
6191 move_task_off_dead_cpu(src_cpu, p);
6192 } while_each_thread(t, p);
6194 read_unlock(&tasklist_lock);
6198 * Schedules idle task to be the next runnable task on current CPU.
6199 * It does so by boosting its priority to highest possible.
6200 * Used by CPU offline code.
6202 void sched_idle_next(void)
6204 int this_cpu = smp_processor_id();
6205 struct rq *rq = cpu_rq(this_cpu);
6206 struct task_struct *p = rq->idle;
6207 unsigned long flags;
6209 /* cpu has to be offline */
6210 BUG_ON(cpu_online(this_cpu));
6213 * Strictly not necessary since rest of the CPUs are stopped by now
6214 * and interrupts disabled on the current cpu.
6216 spin_lock_irqsave(&rq->lock, flags);
6218 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6220 update_rq_clock(rq);
6221 activate_task(rq, p, 0);
6223 spin_unlock_irqrestore(&rq->lock, flags);
6227 * Ensures that the idle task is using init_mm right before its cpu goes
6230 void idle_task_exit(void)
6232 struct mm_struct *mm = current->active_mm;
6234 BUG_ON(cpu_online(smp_processor_id()));
6237 switch_mm(mm, &init_mm, current);
6241 /* called under rq->lock with disabled interrupts */
6242 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6244 struct rq *rq = cpu_rq(dead_cpu);
6246 /* Must be exiting, otherwise would be on tasklist. */
6247 BUG_ON(!p->exit_state);
6249 /* Cannot have done final schedule yet: would have vanished. */
6250 BUG_ON(p->state == TASK_DEAD);
6255 * Drop lock around migration; if someone else moves it,
6256 * that's OK. No task can be added to this CPU, so iteration is
6259 spin_unlock_irq(&rq->lock);
6260 move_task_off_dead_cpu(dead_cpu, p);
6261 spin_lock_irq(&rq->lock);
6266 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6267 static void migrate_dead_tasks(unsigned int dead_cpu)
6269 struct rq *rq = cpu_rq(dead_cpu);
6270 struct task_struct *next;
6273 if (!rq->nr_running)
6275 update_rq_clock(rq);
6276 next = pick_next_task(rq, rq->curr);
6279 next->sched_class->put_prev_task(rq, next);
6280 migrate_dead(dead_cpu, next);
6284 #endif /* CONFIG_HOTPLUG_CPU */
6286 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6288 static struct ctl_table sd_ctl_dir[] = {
6290 .procname = "sched_domain",
6296 static struct ctl_table sd_ctl_root[] = {
6298 .ctl_name = CTL_KERN,
6299 .procname = "kernel",
6301 .child = sd_ctl_dir,
6306 static struct ctl_table *sd_alloc_ctl_entry(int n)
6308 struct ctl_table *entry =
6309 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6314 static void sd_free_ctl_entry(struct ctl_table **tablep)
6316 struct ctl_table *entry;
6319 * In the intermediate directories, both the child directory and
6320 * procname are dynamically allocated and could fail but the mode
6321 * will always be set. In the lowest directory the names are
6322 * static strings and all have proc handlers.
6324 for (entry = *tablep; entry->mode; entry++) {
6326 sd_free_ctl_entry(&entry->child);
6327 if (entry->proc_handler == NULL)
6328 kfree(entry->procname);
6336 set_table_entry(struct ctl_table *entry,
6337 const char *procname, void *data, int maxlen,
6338 mode_t mode, proc_handler *proc_handler)
6340 entry->procname = procname;
6342 entry->maxlen = maxlen;
6344 entry->proc_handler = proc_handler;
6347 static struct ctl_table *
6348 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6350 struct ctl_table *table = sd_alloc_ctl_entry(13);
6355 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6356 sizeof(long), 0644, proc_doulongvec_minmax);
6357 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6358 sizeof(long), 0644, proc_doulongvec_minmax);
6359 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6360 sizeof(int), 0644, proc_dointvec_minmax);
6361 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6362 sizeof(int), 0644, proc_dointvec_minmax);
6363 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6364 sizeof(int), 0644, proc_dointvec_minmax);
6365 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6366 sizeof(int), 0644, proc_dointvec_minmax);
6367 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6368 sizeof(int), 0644, proc_dointvec_minmax);
6369 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6370 sizeof(int), 0644, proc_dointvec_minmax);
6371 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6372 sizeof(int), 0644, proc_dointvec_minmax);
6373 set_table_entry(&table[9], "cache_nice_tries",
6374 &sd->cache_nice_tries,
6375 sizeof(int), 0644, proc_dointvec_minmax);
6376 set_table_entry(&table[10], "flags", &sd->flags,
6377 sizeof(int), 0644, proc_dointvec_minmax);
6378 set_table_entry(&table[11], "name", sd->name,
6379 CORENAME_MAX_SIZE, 0444, proc_dostring);
6380 /* &table[12] is terminator */
6385 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6387 struct ctl_table *entry, *table;
6388 struct sched_domain *sd;
6389 int domain_num = 0, i;
6392 for_each_domain(cpu, sd)
6394 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6399 for_each_domain(cpu, sd) {
6400 snprintf(buf, 32, "domain%d", i);
6401 entry->procname = kstrdup(buf, GFP_KERNEL);
6403 entry->child = sd_alloc_ctl_domain_table(sd);
6410 static struct ctl_table_header *sd_sysctl_header;
6411 static void register_sched_domain_sysctl(void)
6413 int i, cpu_num = num_online_cpus();
6414 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6417 WARN_ON(sd_ctl_dir[0].child);
6418 sd_ctl_dir[0].child = entry;
6423 for_each_online_cpu(i) {
6424 snprintf(buf, 32, "cpu%d", i);
6425 entry->procname = kstrdup(buf, GFP_KERNEL);
6427 entry->child = sd_alloc_ctl_cpu_table(i);
6431 WARN_ON(sd_sysctl_header);
6432 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6435 /* may be called multiple times per register */
6436 static void unregister_sched_domain_sysctl(void)
6438 if (sd_sysctl_header)
6439 unregister_sysctl_table(sd_sysctl_header);
6440 sd_sysctl_header = NULL;
6441 if (sd_ctl_dir[0].child)
6442 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6445 static void register_sched_domain_sysctl(void)
6448 static void unregister_sched_domain_sysctl(void)
6453 static void set_rq_online(struct rq *rq)
6456 const struct sched_class *class;
6458 cpu_set(rq->cpu, rq->rd->online);
6461 for_each_class(class) {
6462 if (class->rq_online)
6463 class->rq_online(rq);
6468 static void set_rq_offline(struct rq *rq)
6471 const struct sched_class *class;
6473 for_each_class(class) {
6474 if (class->rq_offline)
6475 class->rq_offline(rq);
6478 cpu_clear(rq->cpu, rq->rd->online);
6484 * migration_call - callback that gets triggered when a CPU is added.
6485 * Here we can start up the necessary migration thread for the new CPU.
6487 static int __cpuinit
6488 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6490 struct task_struct *p;
6491 int cpu = (long)hcpu;
6492 unsigned long flags;
6497 case CPU_UP_PREPARE:
6498 case CPU_UP_PREPARE_FROZEN:
6499 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6502 kthread_bind(p, cpu);
6503 /* Must be high prio: stop_machine expects to yield to it. */
6504 rq = task_rq_lock(p, &flags);
6505 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6506 task_rq_unlock(rq, &flags);
6507 cpu_rq(cpu)->migration_thread = p;
6511 case CPU_ONLINE_FROZEN:
6512 /* Strictly unnecessary, as first user will wake it. */
6513 wake_up_process(cpu_rq(cpu)->migration_thread);
6515 /* Update our root-domain */
6517 spin_lock_irqsave(&rq->lock, flags);
6519 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6523 spin_unlock_irqrestore(&rq->lock, flags);
6526 #ifdef CONFIG_HOTPLUG_CPU
6527 case CPU_UP_CANCELED:
6528 case CPU_UP_CANCELED_FROZEN:
6529 if (!cpu_rq(cpu)->migration_thread)
6531 /* Unbind it from offline cpu so it can run. Fall thru. */
6532 kthread_bind(cpu_rq(cpu)->migration_thread,
6533 any_online_cpu(cpu_online_map));
6534 kthread_stop(cpu_rq(cpu)->migration_thread);
6535 cpu_rq(cpu)->migration_thread = NULL;
6539 case CPU_DEAD_FROZEN:
6540 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6541 migrate_live_tasks(cpu);
6543 kthread_stop(rq->migration_thread);
6544 rq->migration_thread = NULL;
6545 /* Idle task back to normal (off runqueue, low prio) */
6546 spin_lock_irq(&rq->lock);
6547 update_rq_clock(rq);
6548 deactivate_task(rq, rq->idle, 0);
6549 rq->idle->static_prio = MAX_PRIO;
6550 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6551 rq->idle->sched_class = &idle_sched_class;
6552 migrate_dead_tasks(cpu);
6553 spin_unlock_irq(&rq->lock);
6555 migrate_nr_uninterruptible(rq);
6556 BUG_ON(rq->nr_running != 0);
6559 * No need to migrate the tasks: it was best-effort if
6560 * they didn't take sched_hotcpu_mutex. Just wake up
6563 spin_lock_irq(&rq->lock);
6564 while (!list_empty(&rq->migration_queue)) {
6565 struct migration_req *req;
6567 req = list_entry(rq->migration_queue.next,
6568 struct migration_req, list);
6569 list_del_init(&req->list);
6570 complete(&req->done);
6572 spin_unlock_irq(&rq->lock);
6576 case CPU_DYING_FROZEN:
6577 /* Update our root-domain */
6579 spin_lock_irqsave(&rq->lock, flags);
6581 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6584 spin_unlock_irqrestore(&rq->lock, flags);
6591 /* Register at highest priority so that task migration (migrate_all_tasks)
6592 * happens before everything else.
6594 static struct notifier_block __cpuinitdata migration_notifier = {
6595 .notifier_call = migration_call,
6599 static int __init migration_init(void)
6601 void *cpu = (void *)(long)smp_processor_id();
6604 /* Start one for the boot CPU: */
6605 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6606 BUG_ON(err == NOTIFY_BAD);
6607 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6608 register_cpu_notifier(&migration_notifier);
6612 early_initcall(migration_init);
6617 #ifdef CONFIG_SCHED_DEBUG
6619 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6632 case SD_LV_ALLNODES:
6641 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6642 cpumask_t *groupmask)
6644 struct sched_group *group = sd->groups;
6647 cpulist_scnprintf(str, sizeof(str), sd->span);
6648 cpus_clear(*groupmask);
6650 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6652 if (!(sd->flags & SD_LOAD_BALANCE)) {
6653 printk("does not load-balance\n");
6655 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6660 printk(KERN_CONT "span %s level %s\n",
6661 str, sd_level_to_string(sd->level));
6663 if (!cpu_isset(cpu, sd->span)) {
6664 printk(KERN_ERR "ERROR: domain->span does not contain "
6667 if (!cpu_isset(cpu, group->cpumask)) {
6668 printk(KERN_ERR "ERROR: domain->groups does not contain"
6672 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6676 printk(KERN_ERR "ERROR: group is NULL\n");
6680 if (!group->__cpu_power) {
6681 printk(KERN_CONT "\n");
6682 printk(KERN_ERR "ERROR: domain->cpu_power not "
6687 if (!cpus_weight(group->cpumask)) {
6688 printk(KERN_CONT "\n");
6689 printk(KERN_ERR "ERROR: empty group\n");
6693 if (cpus_intersects(*groupmask, group->cpumask)) {
6694 printk(KERN_CONT "\n");
6695 printk(KERN_ERR "ERROR: repeated CPUs\n");
6699 cpus_or(*groupmask, *groupmask, group->cpumask);
6701 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6702 printk(KERN_CONT " %s", str);
6704 group = group->next;
6705 } while (group != sd->groups);
6706 printk(KERN_CONT "\n");
6708 if (!cpus_equal(sd->span, *groupmask))
6709 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6711 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6712 printk(KERN_ERR "ERROR: parent span is not a superset "
6713 "of domain->span\n");
6717 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6719 cpumask_t *groupmask;
6723 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6727 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6729 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6731 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6736 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6745 #else /* !CONFIG_SCHED_DEBUG */
6746 # define sched_domain_debug(sd, cpu) do { } while (0)
6747 #endif /* CONFIG_SCHED_DEBUG */
6749 static int sd_degenerate(struct sched_domain *sd)
6751 if (cpus_weight(sd->span) == 1)
6754 /* Following flags need at least 2 groups */
6755 if (sd->flags & (SD_LOAD_BALANCE |
6756 SD_BALANCE_NEWIDLE |
6760 SD_SHARE_PKG_RESOURCES)) {
6761 if (sd->groups != sd->groups->next)
6765 /* Following flags don't use groups */
6766 if (sd->flags & (SD_WAKE_IDLE |
6775 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6777 unsigned long cflags = sd->flags, pflags = parent->flags;
6779 if (sd_degenerate(parent))
6782 if (!cpus_equal(sd->span, parent->span))
6785 /* Does parent contain flags not in child? */
6786 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6787 if (cflags & SD_WAKE_AFFINE)
6788 pflags &= ~SD_WAKE_BALANCE;
6789 /* Flags needing groups don't count if only 1 group in parent */
6790 if (parent->groups == parent->groups->next) {
6791 pflags &= ~(SD_LOAD_BALANCE |
6792 SD_BALANCE_NEWIDLE |
6796 SD_SHARE_PKG_RESOURCES);
6798 if (~cflags & pflags)
6804 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6806 unsigned long flags;
6808 spin_lock_irqsave(&rq->lock, flags);
6811 struct root_domain *old_rd = rq->rd;
6813 if (cpu_isset(rq->cpu, old_rd->online))
6816 cpu_clear(rq->cpu, old_rd->span);
6818 if (atomic_dec_and_test(&old_rd->refcount))
6822 atomic_inc(&rd->refcount);
6825 cpu_set(rq->cpu, rd->span);
6826 if (cpu_isset(rq->cpu, cpu_online_map))
6829 spin_unlock_irqrestore(&rq->lock, flags);
6832 static void init_rootdomain(struct root_domain *rd)
6834 memset(rd, 0, sizeof(*rd));
6836 cpus_clear(rd->span);
6837 cpus_clear(rd->online);
6839 cpupri_init(&rd->cpupri);
6842 static void init_defrootdomain(void)
6844 init_rootdomain(&def_root_domain);
6845 atomic_set(&def_root_domain.refcount, 1);
6848 static struct root_domain *alloc_rootdomain(void)
6850 struct root_domain *rd;
6852 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6856 init_rootdomain(rd);
6862 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6863 * hold the hotplug lock.
6866 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6868 struct rq *rq = cpu_rq(cpu);
6869 struct sched_domain *tmp;
6871 /* Remove the sched domains which do not contribute to scheduling. */
6872 for (tmp = sd; tmp; tmp = tmp->parent) {
6873 struct sched_domain *parent = tmp->parent;
6876 if (sd_parent_degenerate(tmp, parent)) {
6877 tmp->parent = parent->parent;
6879 parent->parent->child = tmp;
6883 if (sd && sd_degenerate(sd)) {
6889 sched_domain_debug(sd, cpu);
6891 rq_attach_root(rq, rd);
6892 rcu_assign_pointer(rq->sd, sd);
6895 /* cpus with isolated domains */
6896 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6898 /* Setup the mask of cpus configured for isolated domains */
6899 static int __init isolated_cpu_setup(char *str)
6901 static int __initdata ints[NR_CPUS];
6904 str = get_options(str, ARRAY_SIZE(ints), ints);
6905 cpus_clear(cpu_isolated_map);
6906 for (i = 1; i <= ints[0]; i++)
6907 if (ints[i] < NR_CPUS)
6908 cpu_set(ints[i], cpu_isolated_map);
6912 __setup("isolcpus=", isolated_cpu_setup);
6915 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6916 * to a function which identifies what group(along with sched group) a CPU
6917 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6918 * (due to the fact that we keep track of groups covered with a cpumask_t).
6920 * init_sched_build_groups will build a circular linked list of the groups
6921 * covered by the given span, and will set each group's ->cpumask correctly,
6922 * and ->cpu_power to 0.
6925 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6926 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6927 struct sched_group **sg,
6928 cpumask_t *tmpmask),
6929 cpumask_t *covered, cpumask_t *tmpmask)
6931 struct sched_group *first = NULL, *last = NULL;
6934 cpus_clear(*covered);
6936 for_each_cpu_mask_nr(i, *span) {
6937 struct sched_group *sg;
6938 int group = group_fn(i, cpu_map, &sg, tmpmask);
6941 if (cpu_isset(i, *covered))
6944 cpus_clear(sg->cpumask);
6945 sg->__cpu_power = 0;
6947 for_each_cpu_mask_nr(j, *span) {
6948 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6951 cpu_set(j, *covered);
6952 cpu_set(j, sg->cpumask);
6963 #define SD_NODES_PER_DOMAIN 16
6968 * find_next_best_node - find the next node to include in a sched_domain
6969 * @node: node whose sched_domain we're building
6970 * @used_nodes: nodes already in the sched_domain
6972 * Find the next node to include in a given scheduling domain. Simply
6973 * finds the closest node not already in the @used_nodes map.
6975 * Should use nodemask_t.
6977 static int find_next_best_node(int node, nodemask_t *used_nodes)
6979 int i, n, val, min_val, best_node = 0;
6983 for (i = 0; i < nr_node_ids; i++) {
6984 /* Start at @node */
6985 n = (node + i) % nr_node_ids;
6987 if (!nr_cpus_node(n))
6990 /* Skip already used nodes */
6991 if (node_isset(n, *used_nodes))
6994 /* Simple min distance search */
6995 val = node_distance(node, n);
6997 if (val < min_val) {
7003 node_set(best_node, *used_nodes);
7008 * sched_domain_node_span - get a cpumask for a node's sched_domain
7009 * @node: node whose cpumask we're constructing
7010 * @span: resulting cpumask
7012 * Given a node, construct a good cpumask for its sched_domain to span. It
7013 * should be one that prevents unnecessary balancing, but also spreads tasks
7016 static void sched_domain_node_span(int node, cpumask_t *span)
7018 nodemask_t used_nodes;
7019 node_to_cpumask_ptr(nodemask, node);
7023 nodes_clear(used_nodes);
7025 cpus_or(*span, *span, *nodemask);
7026 node_set(node, used_nodes);
7028 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7029 int next_node = find_next_best_node(node, &used_nodes);
7031 node_to_cpumask_ptr_next(nodemask, next_node);
7032 cpus_or(*span, *span, *nodemask);
7035 #endif /* CONFIG_NUMA */
7037 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7040 * SMT sched-domains:
7042 #ifdef CONFIG_SCHED_SMT
7043 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
7044 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
7047 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7051 *sg = &per_cpu(sched_group_cpus, cpu);
7054 #endif /* CONFIG_SCHED_SMT */
7057 * multi-core sched-domains:
7059 #ifdef CONFIG_SCHED_MC
7060 static DEFINE_PER_CPU(struct sched_domain, core_domains);
7061 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
7062 #endif /* CONFIG_SCHED_MC */
7064 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7066 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7071 *mask = per_cpu(cpu_sibling_map, cpu);
7072 cpus_and(*mask, *mask, *cpu_map);
7073 group = first_cpu(*mask);
7075 *sg = &per_cpu(sched_group_core, group);
7078 #elif defined(CONFIG_SCHED_MC)
7080 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7084 *sg = &per_cpu(sched_group_core, cpu);
7089 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
7090 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
7093 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7097 #ifdef CONFIG_SCHED_MC
7098 *mask = cpu_coregroup_map(cpu);
7099 cpus_and(*mask, *mask, *cpu_map);
7100 group = first_cpu(*mask);
7101 #elif defined(CONFIG_SCHED_SMT)
7102 *mask = per_cpu(cpu_sibling_map, cpu);
7103 cpus_and(*mask, *mask, *cpu_map);
7104 group = first_cpu(*mask);
7109 *sg = &per_cpu(sched_group_phys, group);
7115 * The init_sched_build_groups can't handle what we want to do with node
7116 * groups, so roll our own. Now each node has its own list of groups which
7117 * gets dynamically allocated.
7119 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7120 static struct sched_group ***sched_group_nodes_bycpu;
7122 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7123 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7125 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7126 struct sched_group **sg, cpumask_t *nodemask)
7130 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7131 cpus_and(*nodemask, *nodemask, *cpu_map);
7132 group = first_cpu(*nodemask);
7135 *sg = &per_cpu(sched_group_allnodes, group);
7139 static void init_numa_sched_groups_power(struct sched_group *group_head)
7141 struct sched_group *sg = group_head;
7147 for_each_cpu_mask_nr(j, sg->cpumask) {
7148 struct sched_domain *sd;
7150 sd = &per_cpu(phys_domains, j);
7151 if (j != first_cpu(sd->groups->cpumask)) {
7153 * Only add "power" once for each
7159 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7162 } while (sg != group_head);
7164 #endif /* CONFIG_NUMA */
7167 /* Free memory allocated for various sched_group structures */
7168 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7172 for_each_cpu_mask_nr(cpu, *cpu_map) {
7173 struct sched_group **sched_group_nodes
7174 = sched_group_nodes_bycpu[cpu];
7176 if (!sched_group_nodes)
7179 for (i = 0; i < nr_node_ids; i++) {
7180 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7182 *nodemask = node_to_cpumask(i);
7183 cpus_and(*nodemask, *nodemask, *cpu_map);
7184 if (cpus_empty(*nodemask))
7194 if (oldsg != sched_group_nodes[i])
7197 kfree(sched_group_nodes);
7198 sched_group_nodes_bycpu[cpu] = NULL;
7201 #else /* !CONFIG_NUMA */
7202 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7205 #endif /* CONFIG_NUMA */
7208 * Initialize sched groups cpu_power.
7210 * cpu_power indicates the capacity of sched group, which is used while
7211 * distributing the load between different sched groups in a sched domain.
7212 * Typically cpu_power for all the groups in a sched domain will be same unless
7213 * there are asymmetries in the topology. If there are asymmetries, group
7214 * having more cpu_power will pickup more load compared to the group having
7217 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7218 * the maximum number of tasks a group can handle in the presence of other idle
7219 * or lightly loaded groups in the same sched domain.
7221 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7223 struct sched_domain *child;
7224 struct sched_group *group;
7226 WARN_ON(!sd || !sd->groups);
7228 if (cpu != first_cpu(sd->groups->cpumask))
7233 sd->groups->__cpu_power = 0;
7236 * For perf policy, if the groups in child domain share resources
7237 * (for example cores sharing some portions of the cache hierarchy
7238 * or SMT), then set this domain groups cpu_power such that each group
7239 * can handle only one task, when there are other idle groups in the
7240 * same sched domain.
7242 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7244 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7245 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7250 * add cpu_power of each child group to this groups cpu_power
7252 group = child->groups;
7254 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7255 group = group->next;
7256 } while (group != child->groups);
7260 * Initializers for schedule domains
7261 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7264 #ifdef CONFIG_SCHED_DEBUG
7265 # define SD_INIT_NAME(sd, type) sd->name = #type
7267 # define SD_INIT_NAME(sd, type) do { } while (0)
7270 #define SD_INIT(sd, type) sd_init_##type(sd)
7272 #define SD_INIT_FUNC(type) \
7273 static noinline void sd_init_##type(struct sched_domain *sd) \
7275 memset(sd, 0, sizeof(*sd)); \
7276 *sd = SD_##type##_INIT; \
7277 sd->level = SD_LV_##type; \
7278 SD_INIT_NAME(sd, type); \
7283 SD_INIT_FUNC(ALLNODES)
7286 #ifdef CONFIG_SCHED_SMT
7287 SD_INIT_FUNC(SIBLING)
7289 #ifdef CONFIG_SCHED_MC
7294 * To minimize stack usage kmalloc room for cpumasks and share the
7295 * space as the usage in build_sched_domains() dictates. Used only
7296 * if the amount of space is significant.
7299 cpumask_t tmpmask; /* make this one first */
7302 cpumask_t this_sibling_map;
7303 cpumask_t this_core_map;
7305 cpumask_t send_covered;
7308 cpumask_t domainspan;
7310 cpumask_t notcovered;
7315 #define SCHED_CPUMASK_ALLOC 1
7316 #define SCHED_CPUMASK_FREE(v) kfree(v)
7317 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7319 #define SCHED_CPUMASK_ALLOC 0
7320 #define SCHED_CPUMASK_FREE(v)
7321 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7324 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7325 ((unsigned long)(a) + offsetof(struct allmasks, v))
7327 static int default_relax_domain_level = -1;
7329 static int __init setup_relax_domain_level(char *str)
7333 val = simple_strtoul(str, NULL, 0);
7334 if (val < SD_LV_MAX)
7335 default_relax_domain_level = val;
7339 __setup("relax_domain_level=", setup_relax_domain_level);
7341 static void set_domain_attribute(struct sched_domain *sd,
7342 struct sched_domain_attr *attr)
7346 if (!attr || attr->relax_domain_level < 0) {
7347 if (default_relax_domain_level < 0)
7350 request = default_relax_domain_level;
7352 request = attr->relax_domain_level;
7353 if (request < sd->level) {
7354 /* turn off idle balance on this domain */
7355 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7357 /* turn on idle balance on this domain */
7358 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7363 * Build sched domains for a given set of cpus and attach the sched domains
7364 * to the individual cpus
7366 static int __build_sched_domains(const cpumask_t *cpu_map,
7367 struct sched_domain_attr *attr)
7370 struct root_domain *rd;
7371 SCHED_CPUMASK_DECLARE(allmasks);
7374 struct sched_group **sched_group_nodes = NULL;
7375 int sd_allnodes = 0;
7378 * Allocate the per-node list of sched groups
7380 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7382 if (!sched_group_nodes) {
7383 printk(KERN_WARNING "Can not alloc sched group node list\n");
7388 rd = alloc_rootdomain();
7390 printk(KERN_WARNING "Cannot alloc root domain\n");
7392 kfree(sched_group_nodes);
7397 #if SCHED_CPUMASK_ALLOC
7398 /* get space for all scratch cpumask variables */
7399 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7401 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7404 kfree(sched_group_nodes);
7409 tmpmask = (cpumask_t *)allmasks;
7413 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7417 * Set up domains for cpus specified by the cpu_map.
7419 for_each_cpu_mask_nr(i, *cpu_map) {
7420 struct sched_domain *sd = NULL, *p;
7421 SCHED_CPUMASK_VAR(nodemask, allmasks);
7423 *nodemask = node_to_cpumask(cpu_to_node(i));
7424 cpus_and(*nodemask, *nodemask, *cpu_map);
7427 if (cpus_weight(*cpu_map) >
7428 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7429 sd = &per_cpu(allnodes_domains, i);
7430 SD_INIT(sd, ALLNODES);
7431 set_domain_attribute(sd, attr);
7432 sd->span = *cpu_map;
7433 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7439 sd = &per_cpu(node_domains, i);
7441 set_domain_attribute(sd, attr);
7442 sched_domain_node_span(cpu_to_node(i), &sd->span);
7446 cpus_and(sd->span, sd->span, *cpu_map);
7450 sd = &per_cpu(phys_domains, i);
7452 set_domain_attribute(sd, attr);
7453 sd->span = *nodemask;
7457 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7459 #ifdef CONFIG_SCHED_MC
7461 sd = &per_cpu(core_domains, i);
7463 set_domain_attribute(sd, attr);
7464 sd->span = cpu_coregroup_map(i);
7465 cpus_and(sd->span, sd->span, *cpu_map);
7468 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7471 #ifdef CONFIG_SCHED_SMT
7473 sd = &per_cpu(cpu_domains, i);
7474 SD_INIT(sd, SIBLING);
7475 set_domain_attribute(sd, attr);
7476 sd->span = per_cpu(cpu_sibling_map, i);
7477 cpus_and(sd->span, sd->span, *cpu_map);
7480 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7484 #ifdef CONFIG_SCHED_SMT
7485 /* Set up CPU (sibling) groups */
7486 for_each_cpu_mask_nr(i, *cpu_map) {
7487 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7488 SCHED_CPUMASK_VAR(send_covered, allmasks);
7490 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7491 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7492 if (i != first_cpu(*this_sibling_map))
7495 init_sched_build_groups(this_sibling_map, cpu_map,
7497 send_covered, tmpmask);
7501 #ifdef CONFIG_SCHED_MC
7502 /* Set up multi-core groups */
7503 for_each_cpu_mask_nr(i, *cpu_map) {
7504 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7505 SCHED_CPUMASK_VAR(send_covered, allmasks);
7507 *this_core_map = cpu_coregroup_map(i);
7508 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7509 if (i != first_cpu(*this_core_map))
7512 init_sched_build_groups(this_core_map, cpu_map,
7514 send_covered, tmpmask);
7518 /* Set up physical groups */
7519 for (i = 0; i < nr_node_ids; i++) {
7520 SCHED_CPUMASK_VAR(nodemask, allmasks);
7521 SCHED_CPUMASK_VAR(send_covered, allmasks);
7523 *nodemask = node_to_cpumask(i);
7524 cpus_and(*nodemask, *nodemask, *cpu_map);
7525 if (cpus_empty(*nodemask))
7528 init_sched_build_groups(nodemask, cpu_map,
7530 send_covered, tmpmask);
7534 /* Set up node groups */
7536 SCHED_CPUMASK_VAR(send_covered, allmasks);
7538 init_sched_build_groups(cpu_map, cpu_map,
7539 &cpu_to_allnodes_group,
7540 send_covered, tmpmask);
7543 for (i = 0; i < nr_node_ids; i++) {
7544 /* Set up node groups */
7545 struct sched_group *sg, *prev;
7546 SCHED_CPUMASK_VAR(nodemask, allmasks);
7547 SCHED_CPUMASK_VAR(domainspan, allmasks);
7548 SCHED_CPUMASK_VAR(covered, allmasks);
7551 *nodemask = node_to_cpumask(i);
7552 cpus_clear(*covered);
7554 cpus_and(*nodemask, *nodemask, *cpu_map);
7555 if (cpus_empty(*nodemask)) {
7556 sched_group_nodes[i] = NULL;
7560 sched_domain_node_span(i, domainspan);
7561 cpus_and(*domainspan, *domainspan, *cpu_map);
7563 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7565 printk(KERN_WARNING "Can not alloc domain group for "
7569 sched_group_nodes[i] = sg;
7570 for_each_cpu_mask_nr(j, *nodemask) {
7571 struct sched_domain *sd;
7573 sd = &per_cpu(node_domains, j);
7576 sg->__cpu_power = 0;
7577 sg->cpumask = *nodemask;
7579 cpus_or(*covered, *covered, *nodemask);
7582 for (j = 0; j < nr_node_ids; j++) {
7583 SCHED_CPUMASK_VAR(notcovered, allmasks);
7584 int n = (i + j) % nr_node_ids;
7585 node_to_cpumask_ptr(pnodemask, n);
7587 cpus_complement(*notcovered, *covered);
7588 cpus_and(*tmpmask, *notcovered, *cpu_map);
7589 cpus_and(*tmpmask, *tmpmask, *domainspan);
7590 if (cpus_empty(*tmpmask))
7593 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7594 if (cpus_empty(*tmpmask))
7597 sg = kmalloc_node(sizeof(struct sched_group),
7601 "Can not alloc domain group for node %d\n", j);
7604 sg->__cpu_power = 0;
7605 sg->cpumask = *tmpmask;
7606 sg->next = prev->next;
7607 cpus_or(*covered, *covered, *tmpmask);
7614 /* Calculate CPU power for physical packages and nodes */
7615 #ifdef CONFIG_SCHED_SMT
7616 for_each_cpu_mask_nr(i, *cpu_map) {
7617 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7619 init_sched_groups_power(i, sd);
7622 #ifdef CONFIG_SCHED_MC
7623 for_each_cpu_mask_nr(i, *cpu_map) {
7624 struct sched_domain *sd = &per_cpu(core_domains, i);
7626 init_sched_groups_power(i, sd);
7630 for_each_cpu_mask_nr(i, *cpu_map) {
7631 struct sched_domain *sd = &per_cpu(phys_domains, i);
7633 init_sched_groups_power(i, sd);
7637 for (i = 0; i < nr_node_ids; i++)
7638 init_numa_sched_groups_power(sched_group_nodes[i]);
7641 struct sched_group *sg;
7643 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7645 init_numa_sched_groups_power(sg);
7649 /* Attach the domains */
7650 for_each_cpu_mask_nr(i, *cpu_map) {
7651 struct sched_domain *sd;
7652 #ifdef CONFIG_SCHED_SMT
7653 sd = &per_cpu(cpu_domains, i);
7654 #elif defined(CONFIG_SCHED_MC)
7655 sd = &per_cpu(core_domains, i);
7657 sd = &per_cpu(phys_domains, i);
7659 cpu_attach_domain(sd, rd, i);
7662 SCHED_CPUMASK_FREE((void *)allmasks);
7667 free_sched_groups(cpu_map, tmpmask);
7668 SCHED_CPUMASK_FREE((void *)allmasks);
7673 static int build_sched_domains(const cpumask_t *cpu_map)
7675 return __build_sched_domains(cpu_map, NULL);
7678 static cpumask_t *doms_cur; /* current sched domains */
7679 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7680 static struct sched_domain_attr *dattr_cur;
7681 /* attribues of custom domains in 'doms_cur' */
7684 * Special case: If a kmalloc of a doms_cur partition (array of
7685 * cpumask_t) fails, then fallback to a single sched domain,
7686 * as determined by the single cpumask_t fallback_doms.
7688 static cpumask_t fallback_doms;
7690 void __attribute__((weak)) arch_update_cpu_topology(void)
7695 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7696 * For now this just excludes isolated cpus, but could be used to
7697 * exclude other special cases in the future.
7699 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7703 arch_update_cpu_topology();
7705 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7707 doms_cur = &fallback_doms;
7708 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7710 err = build_sched_domains(doms_cur);
7711 register_sched_domain_sysctl();
7716 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7719 free_sched_groups(cpu_map, tmpmask);
7723 * Detach sched domains from a group of cpus specified in cpu_map
7724 * These cpus will now be attached to the NULL domain
7726 static void detach_destroy_domains(const cpumask_t *cpu_map)
7731 unregister_sched_domain_sysctl();
7733 for_each_cpu_mask_nr(i, *cpu_map)
7734 cpu_attach_domain(NULL, &def_root_domain, i);
7735 synchronize_sched();
7736 arch_destroy_sched_domains(cpu_map, &tmpmask);
7739 /* handle null as "default" */
7740 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7741 struct sched_domain_attr *new, int idx_new)
7743 struct sched_domain_attr tmp;
7750 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7751 new ? (new + idx_new) : &tmp,
7752 sizeof(struct sched_domain_attr));
7756 * Partition sched domains as specified by the 'ndoms_new'
7757 * cpumasks in the array doms_new[] of cpumasks. This compares
7758 * doms_new[] to the current sched domain partitioning, doms_cur[].
7759 * It destroys each deleted domain and builds each new domain.
7761 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7762 * The masks don't intersect (don't overlap.) We should setup one
7763 * sched domain for each mask. CPUs not in any of the cpumasks will
7764 * not be load balanced. If the same cpumask appears both in the
7765 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7768 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7769 * ownership of it and will kfree it when done with it. If the caller
7770 * failed the kmalloc call, then it can pass in doms_new == NULL,
7771 * and partition_sched_domains() will fallback to the single partition
7772 * 'fallback_doms', it also forces the domains to be rebuilt.
7774 * If doms_new==NULL it will be replaced with cpu_online_map.
7775 * ndoms_new==0 is a special case for destroying existing domains.
7776 * It will not create the default domain.
7778 * Call with hotplug lock held
7780 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7781 struct sched_domain_attr *dattr_new)
7785 mutex_lock(&sched_domains_mutex);
7787 /* always unregister in case we don't destroy any domains */
7788 unregister_sched_domain_sysctl();
7790 n = doms_new ? ndoms_new : 0;
7792 /* Destroy deleted domains */
7793 for (i = 0; i < ndoms_cur; i++) {
7794 for (j = 0; j < n; j++) {
7795 if (cpus_equal(doms_cur[i], doms_new[j])
7796 && dattrs_equal(dattr_cur, i, dattr_new, j))
7799 /* no match - a current sched domain not in new doms_new[] */
7800 detach_destroy_domains(doms_cur + i);
7805 if (doms_new == NULL) {
7807 doms_new = &fallback_doms;
7808 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7812 /* Build new domains */
7813 for (i = 0; i < ndoms_new; i++) {
7814 for (j = 0; j < ndoms_cur; j++) {
7815 if (cpus_equal(doms_new[i], doms_cur[j])
7816 && dattrs_equal(dattr_new, i, dattr_cur, j))
7819 /* no match - add a new doms_new */
7820 __build_sched_domains(doms_new + i,
7821 dattr_new ? dattr_new + i : NULL);
7826 /* Remember the new sched domains */
7827 if (doms_cur != &fallback_doms)
7829 kfree(dattr_cur); /* kfree(NULL) is safe */
7830 doms_cur = doms_new;
7831 dattr_cur = dattr_new;
7832 ndoms_cur = ndoms_new;
7834 register_sched_domain_sysctl();
7836 mutex_unlock(&sched_domains_mutex);
7839 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7840 int arch_reinit_sched_domains(void)
7844 /* Destroy domains first to force the rebuild */
7845 partition_sched_domains(0, NULL, NULL);
7847 rebuild_sched_domains();
7853 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7857 if (buf[0] != '0' && buf[0] != '1')
7861 sched_smt_power_savings = (buf[0] == '1');
7863 sched_mc_power_savings = (buf[0] == '1');
7865 ret = arch_reinit_sched_domains();
7867 return ret ? ret : count;
7870 #ifdef CONFIG_SCHED_MC
7871 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7874 return sprintf(page, "%u\n", sched_mc_power_savings);
7876 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7877 const char *buf, size_t count)
7879 return sched_power_savings_store(buf, count, 0);
7881 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7882 sched_mc_power_savings_show,
7883 sched_mc_power_savings_store);
7886 #ifdef CONFIG_SCHED_SMT
7887 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7890 return sprintf(page, "%u\n", sched_smt_power_savings);
7892 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7893 const char *buf, size_t count)
7895 return sched_power_savings_store(buf, count, 1);
7897 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7898 sched_smt_power_savings_show,
7899 sched_smt_power_savings_store);
7902 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7906 #ifdef CONFIG_SCHED_SMT
7908 err = sysfs_create_file(&cls->kset.kobj,
7909 &attr_sched_smt_power_savings.attr);
7911 #ifdef CONFIG_SCHED_MC
7912 if (!err && mc_capable())
7913 err = sysfs_create_file(&cls->kset.kobj,
7914 &attr_sched_mc_power_savings.attr);
7918 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7920 #ifndef CONFIG_CPUSETS
7922 * Add online and remove offline CPUs from the scheduler domains.
7923 * When cpusets are enabled they take over this function.
7925 static int update_sched_domains(struct notifier_block *nfb,
7926 unsigned long action, void *hcpu)
7930 case CPU_ONLINE_FROZEN:
7932 case CPU_DEAD_FROZEN:
7933 partition_sched_domains(1, NULL, NULL);
7942 static int update_runtime(struct notifier_block *nfb,
7943 unsigned long action, void *hcpu)
7945 int cpu = (int)(long)hcpu;
7948 case CPU_DOWN_PREPARE:
7949 case CPU_DOWN_PREPARE_FROZEN:
7950 disable_runtime(cpu_rq(cpu));
7953 case CPU_DOWN_FAILED:
7954 case CPU_DOWN_FAILED_FROZEN:
7956 case CPU_ONLINE_FROZEN:
7957 enable_runtime(cpu_rq(cpu));
7965 void __init sched_init_smp(void)
7967 cpumask_t non_isolated_cpus;
7969 #if defined(CONFIG_NUMA)
7970 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7972 BUG_ON(sched_group_nodes_bycpu == NULL);
7975 mutex_lock(&sched_domains_mutex);
7976 arch_init_sched_domains(&cpu_online_map);
7977 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7978 if (cpus_empty(non_isolated_cpus))
7979 cpu_set(smp_processor_id(), non_isolated_cpus);
7980 mutex_unlock(&sched_domains_mutex);
7983 #ifndef CONFIG_CPUSETS
7984 /* XXX: Theoretical race here - CPU may be hotplugged now */
7985 hotcpu_notifier(update_sched_domains, 0);
7988 /* RT runtime code needs to handle some hotplug events */
7989 hotcpu_notifier(update_runtime, 0);
7993 /* Move init over to a non-isolated CPU */
7994 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7996 sched_init_granularity();
7999 void __init sched_init_smp(void)
8001 sched_init_granularity();
8003 #endif /* CONFIG_SMP */
8005 int in_sched_functions(unsigned long addr)
8007 return in_lock_functions(addr) ||
8008 (addr >= (unsigned long)__sched_text_start
8009 && addr < (unsigned long)__sched_text_end);
8012 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8014 cfs_rq->tasks_timeline = RB_ROOT;
8015 INIT_LIST_HEAD(&cfs_rq->tasks);
8016 #ifdef CONFIG_FAIR_GROUP_SCHED
8019 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8022 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8024 struct rt_prio_array *array;
8027 array = &rt_rq->active;
8028 for (i = 0; i < MAX_RT_PRIO; i++) {
8029 INIT_LIST_HEAD(array->queue + i);
8030 __clear_bit(i, array->bitmap);
8032 /* delimiter for bitsearch: */
8033 __set_bit(MAX_RT_PRIO, array->bitmap);
8035 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8036 rt_rq->highest_prio = MAX_RT_PRIO;
8039 rt_rq->rt_nr_migratory = 0;
8040 rt_rq->overloaded = 0;
8044 rt_rq->rt_throttled = 0;
8045 rt_rq->rt_runtime = 0;
8046 spin_lock_init(&rt_rq->rt_runtime_lock);
8048 #ifdef CONFIG_RT_GROUP_SCHED
8049 rt_rq->rt_nr_boosted = 0;
8054 #ifdef CONFIG_FAIR_GROUP_SCHED
8055 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8056 struct sched_entity *se, int cpu, int add,
8057 struct sched_entity *parent)
8059 struct rq *rq = cpu_rq(cpu);
8060 tg->cfs_rq[cpu] = cfs_rq;
8061 init_cfs_rq(cfs_rq, rq);
8064 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8067 /* se could be NULL for init_task_group */
8072 se->cfs_rq = &rq->cfs;
8074 se->cfs_rq = parent->my_q;
8077 se->load.weight = tg->shares;
8078 se->load.inv_weight = 0;
8079 se->parent = parent;
8083 #ifdef CONFIG_RT_GROUP_SCHED
8084 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8085 struct sched_rt_entity *rt_se, int cpu, int add,
8086 struct sched_rt_entity *parent)
8088 struct rq *rq = cpu_rq(cpu);
8090 tg->rt_rq[cpu] = rt_rq;
8091 init_rt_rq(rt_rq, rq);
8093 rt_rq->rt_se = rt_se;
8094 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8096 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8098 tg->rt_se[cpu] = rt_se;
8103 rt_se->rt_rq = &rq->rt;
8105 rt_se->rt_rq = parent->my_q;
8107 rt_se->my_q = rt_rq;
8108 rt_se->parent = parent;
8109 INIT_LIST_HEAD(&rt_se->run_list);
8113 void __init sched_init(void)
8116 unsigned long alloc_size = 0, ptr;
8118 #ifdef CONFIG_FAIR_GROUP_SCHED
8119 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8121 #ifdef CONFIG_RT_GROUP_SCHED
8122 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8124 #ifdef CONFIG_USER_SCHED
8128 * As sched_init() is called before page_alloc is setup,
8129 * we use alloc_bootmem().
8132 ptr = (unsigned long)alloc_bootmem(alloc_size);
8134 #ifdef CONFIG_FAIR_GROUP_SCHED
8135 init_task_group.se = (struct sched_entity **)ptr;
8136 ptr += nr_cpu_ids * sizeof(void **);
8138 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8139 ptr += nr_cpu_ids * sizeof(void **);
8141 #ifdef CONFIG_USER_SCHED
8142 root_task_group.se = (struct sched_entity **)ptr;
8143 ptr += nr_cpu_ids * sizeof(void **);
8145 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8146 ptr += nr_cpu_ids * sizeof(void **);
8147 #endif /* CONFIG_USER_SCHED */
8148 #endif /* CONFIG_FAIR_GROUP_SCHED */
8149 #ifdef CONFIG_RT_GROUP_SCHED
8150 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8151 ptr += nr_cpu_ids * sizeof(void **);
8153 init_task_group.rt_rq = (struct rt_rq **)ptr;
8154 ptr += nr_cpu_ids * sizeof(void **);
8156 #ifdef CONFIG_USER_SCHED
8157 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8158 ptr += nr_cpu_ids * sizeof(void **);
8160 root_task_group.rt_rq = (struct rt_rq **)ptr;
8161 ptr += nr_cpu_ids * sizeof(void **);
8162 #endif /* CONFIG_USER_SCHED */
8163 #endif /* CONFIG_RT_GROUP_SCHED */
8167 init_defrootdomain();
8170 init_rt_bandwidth(&def_rt_bandwidth,
8171 global_rt_period(), global_rt_runtime());
8173 #ifdef CONFIG_RT_GROUP_SCHED
8174 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8175 global_rt_period(), global_rt_runtime());
8176 #ifdef CONFIG_USER_SCHED
8177 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8178 global_rt_period(), RUNTIME_INF);
8179 #endif /* CONFIG_USER_SCHED */
8180 #endif /* CONFIG_RT_GROUP_SCHED */
8182 #ifdef CONFIG_GROUP_SCHED
8183 list_add(&init_task_group.list, &task_groups);
8184 INIT_LIST_HEAD(&init_task_group.children);
8186 #ifdef CONFIG_USER_SCHED
8187 INIT_LIST_HEAD(&root_task_group.children);
8188 init_task_group.parent = &root_task_group;
8189 list_add(&init_task_group.siblings, &root_task_group.children);
8190 #endif /* CONFIG_USER_SCHED */
8191 #endif /* CONFIG_GROUP_SCHED */
8193 for_each_possible_cpu(i) {
8197 spin_lock_init(&rq->lock);
8199 init_cfs_rq(&rq->cfs, rq);
8200 init_rt_rq(&rq->rt, rq);
8201 #ifdef CONFIG_FAIR_GROUP_SCHED
8202 init_task_group.shares = init_task_group_load;
8203 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8204 #ifdef CONFIG_CGROUP_SCHED
8206 * How much cpu bandwidth does init_task_group get?
8208 * In case of task-groups formed thr' the cgroup filesystem, it
8209 * gets 100% of the cpu resources in the system. This overall
8210 * system cpu resource is divided among the tasks of
8211 * init_task_group and its child task-groups in a fair manner,
8212 * based on each entity's (task or task-group's) weight
8213 * (se->load.weight).
8215 * In other words, if init_task_group has 10 tasks of weight
8216 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8217 * then A0's share of the cpu resource is:
8219 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8221 * We achieve this by letting init_task_group's tasks sit
8222 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8224 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8225 #elif defined CONFIG_USER_SCHED
8226 root_task_group.shares = NICE_0_LOAD;
8227 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8229 * In case of task-groups formed thr' the user id of tasks,
8230 * init_task_group represents tasks belonging to root user.
8231 * Hence it forms a sibling of all subsequent groups formed.
8232 * In this case, init_task_group gets only a fraction of overall
8233 * system cpu resource, based on the weight assigned to root
8234 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8235 * by letting tasks of init_task_group sit in a separate cfs_rq
8236 * (init_cfs_rq) and having one entity represent this group of
8237 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8239 init_tg_cfs_entry(&init_task_group,
8240 &per_cpu(init_cfs_rq, i),
8241 &per_cpu(init_sched_entity, i), i, 1,
8242 root_task_group.se[i]);
8245 #endif /* CONFIG_FAIR_GROUP_SCHED */
8247 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8248 #ifdef CONFIG_RT_GROUP_SCHED
8249 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8250 #ifdef CONFIG_CGROUP_SCHED
8251 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8252 #elif defined CONFIG_USER_SCHED
8253 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8254 init_tg_rt_entry(&init_task_group,
8255 &per_cpu(init_rt_rq, i),
8256 &per_cpu(init_sched_rt_entity, i), i, 1,
8257 root_task_group.rt_se[i]);
8261 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8262 rq->cpu_load[j] = 0;
8266 rq->active_balance = 0;
8267 rq->next_balance = jiffies;
8271 rq->migration_thread = NULL;
8272 INIT_LIST_HEAD(&rq->migration_queue);
8273 rq_attach_root(rq, &def_root_domain);
8276 atomic_set(&rq->nr_iowait, 0);
8279 set_load_weight(&init_task);
8281 #ifdef CONFIG_PREEMPT_NOTIFIERS
8282 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8286 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8289 #ifdef CONFIG_RT_MUTEXES
8290 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8294 * The boot idle thread does lazy MMU switching as well:
8296 atomic_inc(&init_mm.mm_count);
8297 enter_lazy_tlb(&init_mm, current);
8300 * Make us the idle thread. Technically, schedule() should not be
8301 * called from this thread, however somewhere below it might be,
8302 * but because we are the idle thread, we just pick up running again
8303 * when this runqueue becomes "idle".
8305 init_idle(current, smp_processor_id());
8307 * During early bootup we pretend to be a normal task:
8309 current->sched_class = &fair_sched_class;
8311 scheduler_running = 1;
8314 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8315 void __might_sleep(char *file, int line)
8318 static unsigned long prev_jiffy; /* ratelimiting */
8320 if ((!in_atomic() && !irqs_disabled()) ||
8321 system_state != SYSTEM_RUNNING || oops_in_progress)
8323 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8325 prev_jiffy = jiffies;
8328 "BUG: sleeping function called from invalid context at %s:%d\n",
8331 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8332 in_atomic(), irqs_disabled(),
8333 current->pid, current->comm);
8335 debug_show_held_locks(current);
8336 if (irqs_disabled())
8337 print_irqtrace_events(current);
8341 EXPORT_SYMBOL(__might_sleep);
8344 #ifdef CONFIG_MAGIC_SYSRQ
8345 static void normalize_task(struct rq *rq, struct task_struct *p)
8349 update_rq_clock(rq);
8350 on_rq = p->se.on_rq;
8352 deactivate_task(rq, p, 0);
8353 __setscheduler(rq, p, SCHED_NORMAL, 0);
8355 activate_task(rq, p, 0);
8356 resched_task(rq->curr);
8360 void normalize_rt_tasks(void)
8362 struct task_struct *g, *p;
8363 unsigned long flags;
8366 read_lock_irqsave(&tasklist_lock, flags);
8367 do_each_thread(g, p) {
8369 * Only normalize user tasks:
8374 p->se.exec_start = 0;
8375 #ifdef CONFIG_SCHEDSTATS
8376 p->se.wait_start = 0;
8377 p->se.sleep_start = 0;
8378 p->se.block_start = 0;
8383 * Renice negative nice level userspace
8386 if (TASK_NICE(p) < 0 && p->mm)
8387 set_user_nice(p, 0);
8391 spin_lock(&p->pi_lock);
8392 rq = __task_rq_lock(p);
8394 normalize_task(rq, p);
8396 __task_rq_unlock(rq);
8397 spin_unlock(&p->pi_lock);
8398 } while_each_thread(g, p);
8400 read_unlock_irqrestore(&tasklist_lock, flags);
8403 #endif /* CONFIG_MAGIC_SYSRQ */
8407 * These functions are only useful for the IA64 MCA handling.
8409 * They can only be called when the whole system has been
8410 * stopped - every CPU needs to be quiescent, and no scheduling
8411 * activity can take place. Using them for anything else would
8412 * be a serious bug, and as a result, they aren't even visible
8413 * under any other configuration.
8417 * curr_task - return the current task for a given cpu.
8418 * @cpu: the processor in question.
8420 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8422 struct task_struct *curr_task(int cpu)
8424 return cpu_curr(cpu);
8428 * set_curr_task - set the current task for a given cpu.
8429 * @cpu: the processor in question.
8430 * @p: the task pointer to set.
8432 * Description: This function must only be used when non-maskable interrupts
8433 * are serviced on a separate stack. It allows the architecture to switch the
8434 * notion of the current task on a cpu in a non-blocking manner. This function
8435 * must be called with all CPU's synchronized, and interrupts disabled, the
8436 * and caller must save the original value of the current task (see
8437 * curr_task() above) and restore that value before reenabling interrupts and
8438 * re-starting the system.
8440 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8442 void set_curr_task(int cpu, struct task_struct *p)
8449 #ifdef CONFIG_FAIR_GROUP_SCHED
8450 static void free_fair_sched_group(struct task_group *tg)
8454 for_each_possible_cpu(i) {
8456 kfree(tg->cfs_rq[i]);
8466 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8468 struct cfs_rq *cfs_rq;
8469 struct sched_entity *se, *parent_se;
8473 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8476 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8480 tg->shares = NICE_0_LOAD;
8482 for_each_possible_cpu(i) {
8485 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8486 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8490 se = kmalloc_node(sizeof(struct sched_entity),
8491 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8495 parent_se = parent ? parent->se[i] : NULL;
8496 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8505 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8507 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8508 &cpu_rq(cpu)->leaf_cfs_rq_list);
8511 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8513 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8515 #else /* !CONFG_FAIR_GROUP_SCHED */
8516 static inline void free_fair_sched_group(struct task_group *tg)
8521 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8526 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8530 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8533 #endif /* CONFIG_FAIR_GROUP_SCHED */
8535 #ifdef CONFIG_RT_GROUP_SCHED
8536 static void free_rt_sched_group(struct task_group *tg)
8540 destroy_rt_bandwidth(&tg->rt_bandwidth);
8542 for_each_possible_cpu(i) {
8544 kfree(tg->rt_rq[i]);
8546 kfree(tg->rt_se[i]);
8554 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8556 struct rt_rq *rt_rq;
8557 struct sched_rt_entity *rt_se, *parent_se;
8561 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8564 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8568 init_rt_bandwidth(&tg->rt_bandwidth,
8569 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8571 for_each_possible_cpu(i) {
8574 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8575 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8579 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8580 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8584 parent_se = parent ? parent->rt_se[i] : NULL;
8585 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8594 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8596 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8597 &cpu_rq(cpu)->leaf_rt_rq_list);
8600 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8602 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8604 #else /* !CONFIG_RT_GROUP_SCHED */
8605 static inline void free_rt_sched_group(struct task_group *tg)
8610 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8615 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8619 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8622 #endif /* CONFIG_RT_GROUP_SCHED */
8624 #ifdef CONFIG_GROUP_SCHED
8625 static void free_sched_group(struct task_group *tg)
8627 free_fair_sched_group(tg);
8628 free_rt_sched_group(tg);
8632 /* allocate runqueue etc for a new task group */
8633 struct task_group *sched_create_group(struct task_group *parent)
8635 struct task_group *tg;
8636 unsigned long flags;
8639 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8641 return ERR_PTR(-ENOMEM);
8643 if (!alloc_fair_sched_group(tg, parent))
8646 if (!alloc_rt_sched_group(tg, parent))
8649 spin_lock_irqsave(&task_group_lock, flags);
8650 for_each_possible_cpu(i) {
8651 register_fair_sched_group(tg, i);
8652 register_rt_sched_group(tg, i);
8654 list_add_rcu(&tg->list, &task_groups);
8656 WARN_ON(!parent); /* root should already exist */
8658 tg->parent = parent;
8659 INIT_LIST_HEAD(&tg->children);
8660 list_add_rcu(&tg->siblings, &parent->children);
8661 spin_unlock_irqrestore(&task_group_lock, flags);
8666 free_sched_group(tg);
8667 return ERR_PTR(-ENOMEM);
8670 /* rcu callback to free various structures associated with a task group */
8671 static void free_sched_group_rcu(struct rcu_head *rhp)
8673 /* now it should be safe to free those cfs_rqs */
8674 free_sched_group(container_of(rhp, struct task_group, rcu));
8677 /* Destroy runqueue etc associated with a task group */
8678 void sched_destroy_group(struct task_group *tg)
8680 unsigned long flags;
8683 spin_lock_irqsave(&task_group_lock, flags);
8684 for_each_possible_cpu(i) {
8685 unregister_fair_sched_group(tg, i);
8686 unregister_rt_sched_group(tg, i);
8688 list_del_rcu(&tg->list);
8689 list_del_rcu(&tg->siblings);
8690 spin_unlock_irqrestore(&task_group_lock, flags);
8692 /* wait for possible concurrent references to cfs_rqs complete */
8693 call_rcu(&tg->rcu, free_sched_group_rcu);
8696 /* change task's runqueue when it moves between groups.
8697 * The caller of this function should have put the task in its new group
8698 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8699 * reflect its new group.
8701 void sched_move_task(struct task_struct *tsk)
8704 unsigned long flags;
8707 rq = task_rq_lock(tsk, &flags);
8709 update_rq_clock(rq);
8711 running = task_current(rq, tsk);
8712 on_rq = tsk->se.on_rq;
8715 dequeue_task(rq, tsk, 0);
8716 if (unlikely(running))
8717 tsk->sched_class->put_prev_task(rq, tsk);
8719 set_task_rq(tsk, task_cpu(tsk));
8721 #ifdef CONFIG_FAIR_GROUP_SCHED
8722 if (tsk->sched_class->moved_group)
8723 tsk->sched_class->moved_group(tsk);
8726 if (unlikely(running))
8727 tsk->sched_class->set_curr_task(rq);
8729 enqueue_task(rq, tsk, 0);
8731 task_rq_unlock(rq, &flags);
8733 #endif /* CONFIG_GROUP_SCHED */
8735 #ifdef CONFIG_FAIR_GROUP_SCHED
8736 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8738 struct cfs_rq *cfs_rq = se->cfs_rq;
8743 dequeue_entity(cfs_rq, se, 0);
8745 se->load.weight = shares;
8746 se->load.inv_weight = 0;
8749 enqueue_entity(cfs_rq, se, 0);
8752 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8754 struct cfs_rq *cfs_rq = se->cfs_rq;
8755 struct rq *rq = cfs_rq->rq;
8756 unsigned long flags;
8758 spin_lock_irqsave(&rq->lock, flags);
8759 __set_se_shares(se, shares);
8760 spin_unlock_irqrestore(&rq->lock, flags);
8763 static DEFINE_MUTEX(shares_mutex);
8765 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8768 unsigned long flags;
8771 * We can't change the weight of the root cgroup.
8776 if (shares < MIN_SHARES)
8777 shares = MIN_SHARES;
8778 else if (shares > MAX_SHARES)
8779 shares = MAX_SHARES;
8781 mutex_lock(&shares_mutex);
8782 if (tg->shares == shares)
8785 spin_lock_irqsave(&task_group_lock, flags);
8786 for_each_possible_cpu(i)
8787 unregister_fair_sched_group(tg, i);
8788 list_del_rcu(&tg->siblings);
8789 spin_unlock_irqrestore(&task_group_lock, flags);
8791 /* wait for any ongoing reference to this group to finish */
8792 synchronize_sched();
8795 * Now we are free to modify the group's share on each cpu
8796 * w/o tripping rebalance_share or load_balance_fair.
8798 tg->shares = shares;
8799 for_each_possible_cpu(i) {
8803 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8804 set_se_shares(tg->se[i], shares);
8808 * Enable load balance activity on this group, by inserting it back on
8809 * each cpu's rq->leaf_cfs_rq_list.
8811 spin_lock_irqsave(&task_group_lock, flags);
8812 for_each_possible_cpu(i)
8813 register_fair_sched_group(tg, i);
8814 list_add_rcu(&tg->siblings, &tg->parent->children);
8815 spin_unlock_irqrestore(&task_group_lock, flags);
8817 mutex_unlock(&shares_mutex);
8821 unsigned long sched_group_shares(struct task_group *tg)
8827 #ifdef CONFIG_RT_GROUP_SCHED
8829 * Ensure that the real time constraints are schedulable.
8831 static DEFINE_MUTEX(rt_constraints_mutex);
8833 static unsigned long to_ratio(u64 period, u64 runtime)
8835 if (runtime == RUNTIME_INF)
8838 return div64_u64(runtime << 20, period);
8841 /* Must be called with tasklist_lock held */
8842 static inline int tg_has_rt_tasks(struct task_group *tg)
8844 struct task_struct *g, *p;
8846 do_each_thread(g, p) {
8847 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8849 } while_each_thread(g, p);
8854 struct rt_schedulable_data {
8855 struct task_group *tg;
8860 static int tg_schedulable(struct task_group *tg, void *data)
8862 struct rt_schedulable_data *d = data;
8863 struct task_group *child;
8864 unsigned long total, sum = 0;
8865 u64 period, runtime;
8867 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8868 runtime = tg->rt_bandwidth.rt_runtime;
8871 period = d->rt_period;
8872 runtime = d->rt_runtime;
8876 * Cannot have more runtime than the period.
8878 if (runtime > period && runtime != RUNTIME_INF)
8882 * Ensure we don't starve existing RT tasks.
8884 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8887 total = to_ratio(period, runtime);
8890 * Nobody can have more than the global setting allows.
8892 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8896 * The sum of our children's runtime should not exceed our own.
8898 list_for_each_entry_rcu(child, &tg->children, siblings) {
8899 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8900 runtime = child->rt_bandwidth.rt_runtime;
8902 if (child == d->tg) {
8903 period = d->rt_period;
8904 runtime = d->rt_runtime;
8907 sum += to_ratio(period, runtime);
8916 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8918 struct rt_schedulable_data data = {
8920 .rt_period = period,
8921 .rt_runtime = runtime,
8924 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8927 static int tg_set_bandwidth(struct task_group *tg,
8928 u64 rt_period, u64 rt_runtime)
8932 mutex_lock(&rt_constraints_mutex);
8933 read_lock(&tasklist_lock);
8934 err = __rt_schedulable(tg, rt_period, rt_runtime);
8938 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8939 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8940 tg->rt_bandwidth.rt_runtime = rt_runtime;
8942 for_each_possible_cpu(i) {
8943 struct rt_rq *rt_rq = tg->rt_rq[i];
8945 spin_lock(&rt_rq->rt_runtime_lock);
8946 rt_rq->rt_runtime = rt_runtime;
8947 spin_unlock(&rt_rq->rt_runtime_lock);
8949 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8951 read_unlock(&tasklist_lock);
8952 mutex_unlock(&rt_constraints_mutex);
8957 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8959 u64 rt_runtime, rt_period;
8961 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8962 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8963 if (rt_runtime_us < 0)
8964 rt_runtime = RUNTIME_INF;
8966 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8969 long sched_group_rt_runtime(struct task_group *tg)
8973 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8976 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8977 do_div(rt_runtime_us, NSEC_PER_USEC);
8978 return rt_runtime_us;
8981 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8983 u64 rt_runtime, rt_period;
8985 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8986 rt_runtime = tg->rt_bandwidth.rt_runtime;
8991 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8994 long sched_group_rt_period(struct task_group *tg)
8998 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8999 do_div(rt_period_us, NSEC_PER_USEC);
9000 return rt_period_us;
9003 static int sched_rt_global_constraints(void)
9005 u64 runtime, period;
9008 if (sysctl_sched_rt_period <= 0)
9011 runtime = global_rt_runtime();
9012 period = global_rt_period();
9015 * Sanity check on the sysctl variables.
9017 if (runtime > period && runtime != RUNTIME_INF)
9020 mutex_lock(&rt_constraints_mutex);
9021 read_lock(&tasklist_lock);
9022 ret = __rt_schedulable(NULL, 0, 0);
9023 read_unlock(&tasklist_lock);
9024 mutex_unlock(&rt_constraints_mutex);
9028 #else /* !CONFIG_RT_GROUP_SCHED */
9029 static int sched_rt_global_constraints(void)
9031 unsigned long flags;
9034 if (sysctl_sched_rt_period <= 0)
9037 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9038 for_each_possible_cpu(i) {
9039 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9041 spin_lock(&rt_rq->rt_runtime_lock);
9042 rt_rq->rt_runtime = global_rt_runtime();
9043 spin_unlock(&rt_rq->rt_runtime_lock);
9045 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9049 #endif /* CONFIG_RT_GROUP_SCHED */
9051 int sched_rt_handler(struct ctl_table *table, int write,
9052 struct file *filp, void __user *buffer, size_t *lenp,
9056 int old_period, old_runtime;
9057 static DEFINE_MUTEX(mutex);
9060 old_period = sysctl_sched_rt_period;
9061 old_runtime = sysctl_sched_rt_runtime;
9063 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9065 if (!ret && write) {
9066 ret = sched_rt_global_constraints();
9068 sysctl_sched_rt_period = old_period;
9069 sysctl_sched_rt_runtime = old_runtime;
9071 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9072 def_rt_bandwidth.rt_period =
9073 ns_to_ktime(global_rt_period());
9076 mutex_unlock(&mutex);
9081 #ifdef CONFIG_CGROUP_SCHED
9083 /* return corresponding task_group object of a cgroup */
9084 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9086 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9087 struct task_group, css);
9090 static struct cgroup_subsys_state *
9091 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9093 struct task_group *tg, *parent;
9095 if (!cgrp->parent) {
9096 /* This is early initialization for the top cgroup */
9097 return &init_task_group.css;
9100 parent = cgroup_tg(cgrp->parent);
9101 tg = sched_create_group(parent);
9103 return ERR_PTR(-ENOMEM);
9109 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9111 struct task_group *tg = cgroup_tg(cgrp);
9113 sched_destroy_group(tg);
9117 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9118 struct task_struct *tsk)
9120 #ifdef CONFIG_RT_GROUP_SCHED
9121 /* Don't accept realtime tasks when there is no way for them to run */
9122 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9125 /* We don't support RT-tasks being in separate groups */
9126 if (tsk->sched_class != &fair_sched_class)
9134 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9135 struct cgroup *old_cont, struct task_struct *tsk)
9137 sched_move_task(tsk);
9140 #ifdef CONFIG_FAIR_GROUP_SCHED
9141 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9144 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9147 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9149 struct task_group *tg = cgroup_tg(cgrp);
9151 return (u64) tg->shares;
9153 #endif /* CONFIG_FAIR_GROUP_SCHED */
9155 #ifdef CONFIG_RT_GROUP_SCHED
9156 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9159 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9162 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9164 return sched_group_rt_runtime(cgroup_tg(cgrp));
9167 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9170 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9173 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9175 return sched_group_rt_period(cgroup_tg(cgrp));
9177 #endif /* CONFIG_RT_GROUP_SCHED */
9179 static struct cftype cpu_files[] = {
9180 #ifdef CONFIG_FAIR_GROUP_SCHED
9183 .read_u64 = cpu_shares_read_u64,
9184 .write_u64 = cpu_shares_write_u64,
9187 #ifdef CONFIG_RT_GROUP_SCHED
9189 .name = "rt_runtime_us",
9190 .read_s64 = cpu_rt_runtime_read,
9191 .write_s64 = cpu_rt_runtime_write,
9194 .name = "rt_period_us",
9195 .read_u64 = cpu_rt_period_read_uint,
9196 .write_u64 = cpu_rt_period_write_uint,
9201 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9203 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9206 struct cgroup_subsys cpu_cgroup_subsys = {
9208 .create = cpu_cgroup_create,
9209 .destroy = cpu_cgroup_destroy,
9210 .can_attach = cpu_cgroup_can_attach,
9211 .attach = cpu_cgroup_attach,
9212 .populate = cpu_cgroup_populate,
9213 .subsys_id = cpu_cgroup_subsys_id,
9217 #endif /* CONFIG_CGROUP_SCHED */
9219 #ifdef CONFIG_CGROUP_CPUACCT
9222 * CPU accounting code for task groups.
9224 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9225 * (balbir@in.ibm.com).
9228 /* track cpu usage of a group of tasks */
9230 struct cgroup_subsys_state css;
9231 /* cpuusage holds pointer to a u64-type object on every cpu */
9235 struct cgroup_subsys cpuacct_subsys;
9237 /* return cpu accounting group corresponding to this container */
9238 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9240 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9241 struct cpuacct, css);
9244 /* return cpu accounting group to which this task belongs */
9245 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9247 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9248 struct cpuacct, css);
9251 /* create a new cpu accounting group */
9252 static struct cgroup_subsys_state *cpuacct_create(
9253 struct cgroup_subsys *ss, struct cgroup *cgrp)
9255 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9258 return ERR_PTR(-ENOMEM);
9260 ca->cpuusage = alloc_percpu(u64);
9261 if (!ca->cpuusage) {
9263 return ERR_PTR(-ENOMEM);
9269 /* destroy an existing cpu accounting group */
9271 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9273 struct cpuacct *ca = cgroup_ca(cgrp);
9275 free_percpu(ca->cpuusage);
9279 /* return total cpu usage (in nanoseconds) of a group */
9280 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9282 struct cpuacct *ca = cgroup_ca(cgrp);
9283 u64 totalcpuusage = 0;
9286 for_each_possible_cpu(i) {
9287 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9290 * Take rq->lock to make 64-bit addition safe on 32-bit
9293 spin_lock_irq(&cpu_rq(i)->lock);
9294 totalcpuusage += *cpuusage;
9295 spin_unlock_irq(&cpu_rq(i)->lock);
9298 return totalcpuusage;
9301 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9304 struct cpuacct *ca = cgroup_ca(cgrp);
9313 for_each_possible_cpu(i) {
9314 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9316 spin_lock_irq(&cpu_rq(i)->lock);
9318 spin_unlock_irq(&cpu_rq(i)->lock);
9324 static struct cftype files[] = {
9327 .read_u64 = cpuusage_read,
9328 .write_u64 = cpuusage_write,
9332 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9334 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9338 * charge this task's execution time to its accounting group.
9340 * called with rq->lock held.
9342 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9346 if (!cpuacct_subsys.active)
9351 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9353 *cpuusage += cputime;
9357 struct cgroup_subsys cpuacct_subsys = {
9359 .create = cpuacct_create,
9360 .destroy = cpuacct_destroy,
9361 .populate = cpuacct_populate,
9362 .subsys_id = cpuacct_subsys_id,
9364 #endif /* CONFIG_CGROUP_CPUACCT */