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>
74 #include <trace/sched.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
82 * Convert user-nice values [ -20 ... 0 ... 19 ]
83 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
86 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
87 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
88 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
91 * 'User priority' is the nice value converted to something we
92 * can work with better when scaling various scheduler parameters,
93 * it's a [ 0 ... 39 ] range.
95 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
96 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
97 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
100 * Helpers for converting nanosecond timing to jiffy resolution
102 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
104 #define NICE_0_LOAD SCHED_LOAD_SCALE
105 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
108 * These are the 'tuning knobs' of the scheduler:
110 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
111 * Timeslices get refilled after they expire.
113 #define DEF_TIMESLICE (100 * HZ / 1000)
116 * single value that denotes runtime == period, ie unlimited time.
118 #define RUNTIME_INF ((u64)~0ULL)
122 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
123 * Since cpu_power is a 'constant', we can use a reciprocal divide.
125 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
127 return reciprocal_divide(load, sg->reciprocal_cpu_power);
131 * Each time a sched group cpu_power is changed,
132 * we must compute its reciprocal value
134 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
136 sg->__cpu_power += val;
137 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
141 static inline int rt_policy(int policy)
143 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
148 static inline int task_has_rt_policy(struct task_struct *p)
150 return rt_policy(p->policy);
154 * This is the priority-queue data structure of the RT scheduling class:
156 struct rt_prio_array {
157 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
158 struct list_head queue[MAX_RT_PRIO];
161 struct rt_bandwidth {
162 /* nests inside the rq lock: */
163 spinlock_t rt_runtime_lock;
166 struct hrtimer rt_period_timer;
169 static struct rt_bandwidth def_rt_bandwidth;
171 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
173 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
175 struct rt_bandwidth *rt_b =
176 container_of(timer, struct rt_bandwidth, rt_period_timer);
182 now = hrtimer_cb_get_time(timer);
183 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
188 idle = do_sched_rt_period_timer(rt_b, overrun);
191 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
195 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
197 rt_b->rt_period = ns_to_ktime(period);
198 rt_b->rt_runtime = runtime;
200 spin_lock_init(&rt_b->rt_runtime_lock);
202 hrtimer_init(&rt_b->rt_period_timer,
203 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
204 rt_b->rt_period_timer.function = sched_rt_period_timer;
205 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_UNLOCKED;
208 static inline int rt_bandwidth_enabled(void)
210 return sysctl_sched_rt_runtime >= 0;
213 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
217 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
220 if (hrtimer_active(&rt_b->rt_period_timer))
223 spin_lock(&rt_b->rt_runtime_lock);
225 if (hrtimer_active(&rt_b->rt_period_timer))
228 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
229 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
230 hrtimer_start_expires(&rt_b->rt_period_timer,
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 hrtimer_set_expires(timer, 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 trace_sched_wait_task(rq, p);
1940 running = task_running(rq, p);
1941 on_rq = p->se.on_rq;
1943 if (!match_state || p->state == match_state)
1944 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1945 task_rq_unlock(rq, &flags);
1948 * If it changed from the expected state, bail out now.
1950 if (unlikely(!ncsw))
1954 * Was it really running after all now that we
1955 * checked with the proper locks actually held?
1957 * Oops. Go back and try again..
1959 if (unlikely(running)) {
1965 * It's not enough that it's not actively running,
1966 * it must be off the runqueue _entirely_, and not
1969 * So if it wa still runnable (but just not actively
1970 * running right now), it's preempted, and we should
1971 * yield - it could be a while.
1973 if (unlikely(on_rq)) {
1974 schedule_timeout_uninterruptible(1);
1979 * Ahh, all good. It wasn't running, and it wasn't
1980 * runnable, which means that it will never become
1981 * running in the future either. We're all done!
1990 * kick_process - kick a running thread to enter/exit the kernel
1991 * @p: the to-be-kicked thread
1993 * Cause a process which is running on another CPU to enter
1994 * kernel-mode, without any delay. (to get signals handled.)
1996 * NOTE: this function doesnt have to take the runqueue lock,
1997 * because all it wants to ensure is that the remote task enters
1998 * the kernel. If the IPI races and the task has been migrated
1999 * to another CPU then no harm is done and the purpose has been
2002 void kick_process(struct task_struct *p)
2008 if ((cpu != smp_processor_id()) && task_curr(p))
2009 smp_send_reschedule(cpu);
2014 * Return a low guess at the load of a migration-source cpu weighted
2015 * according to the scheduling class and "nice" value.
2017 * We want to under-estimate the load of migration sources, to
2018 * balance conservatively.
2020 static unsigned long source_load(int cpu, int type)
2022 struct rq *rq = cpu_rq(cpu);
2023 unsigned long total = weighted_cpuload(cpu);
2025 if (type == 0 || !sched_feat(LB_BIAS))
2028 return min(rq->cpu_load[type-1], total);
2032 * Return a high guess at the load of a migration-target cpu weighted
2033 * according to the scheduling class and "nice" value.
2035 static unsigned long target_load(int cpu, int type)
2037 struct rq *rq = cpu_rq(cpu);
2038 unsigned long total = weighted_cpuload(cpu);
2040 if (type == 0 || !sched_feat(LB_BIAS))
2043 return max(rq->cpu_load[type-1], total);
2047 * find_idlest_group finds and returns the least busy CPU group within the
2050 static struct sched_group *
2051 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2053 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2054 unsigned long min_load = ULONG_MAX, this_load = 0;
2055 int load_idx = sd->forkexec_idx;
2056 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2059 unsigned long load, avg_load;
2063 /* Skip over this group if it has no CPUs allowed */
2064 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2067 local_group = cpu_isset(this_cpu, group->cpumask);
2069 /* Tally up the load of all CPUs in the group */
2072 for_each_cpu_mask_nr(i, group->cpumask) {
2073 /* Bias balancing toward cpus of our domain */
2075 load = source_load(i, load_idx);
2077 load = target_load(i, load_idx);
2082 /* Adjust by relative CPU power of the group */
2083 avg_load = sg_div_cpu_power(group,
2084 avg_load * SCHED_LOAD_SCALE);
2087 this_load = avg_load;
2089 } else if (avg_load < min_load) {
2090 min_load = avg_load;
2093 } while (group = group->next, group != sd->groups);
2095 if (!idlest || 100*this_load < imbalance*min_load)
2101 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2104 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2107 unsigned long load, min_load = ULONG_MAX;
2111 /* Traverse only the allowed CPUs */
2112 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2114 for_each_cpu_mask_nr(i, *tmp) {
2115 load = weighted_cpuload(i);
2117 if (load < min_load || (load == min_load && i == this_cpu)) {
2127 * sched_balance_self: balance the current task (running on cpu) in domains
2128 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2131 * Balance, ie. select the least loaded group.
2133 * Returns the target CPU number, or the same CPU if no balancing is needed.
2135 * preempt must be disabled.
2137 static int sched_balance_self(int cpu, int flag)
2139 struct task_struct *t = current;
2140 struct sched_domain *tmp, *sd = NULL;
2142 for_each_domain(cpu, tmp) {
2144 * If power savings logic is enabled for a domain, stop there.
2146 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2148 if (tmp->flags & flag)
2156 cpumask_t span, tmpmask;
2157 struct sched_group *group;
2158 int new_cpu, weight;
2160 if (!(sd->flags & flag)) {
2166 group = find_idlest_group(sd, t, cpu);
2172 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2173 if (new_cpu == -1 || new_cpu == cpu) {
2174 /* Now try balancing at a lower domain level of cpu */
2179 /* Now try balancing at a lower domain level of new_cpu */
2182 weight = cpus_weight(span);
2183 for_each_domain(cpu, tmp) {
2184 if (weight <= cpus_weight(tmp->span))
2186 if (tmp->flags & flag)
2189 /* while loop will break here if sd == NULL */
2195 #endif /* CONFIG_SMP */
2198 * try_to_wake_up - wake up a thread
2199 * @p: the to-be-woken-up thread
2200 * @state: the mask of task states that can be woken
2201 * @sync: do a synchronous wakeup?
2203 * Put it on the run-queue if it's not already there. The "current"
2204 * thread is always on the run-queue (except when the actual
2205 * re-schedule is in progress), and as such you're allowed to do
2206 * the simpler "current->state = TASK_RUNNING" to mark yourself
2207 * runnable without the overhead of this.
2209 * returns failure only if the task is already active.
2211 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2213 int cpu, orig_cpu, this_cpu, success = 0;
2214 unsigned long flags;
2218 if (!sched_feat(SYNC_WAKEUPS))
2222 if (sched_feat(LB_WAKEUP_UPDATE)) {
2223 struct sched_domain *sd;
2225 this_cpu = raw_smp_processor_id();
2228 for_each_domain(this_cpu, sd) {
2229 if (cpu_isset(cpu, sd->span)) {
2238 rq = task_rq_lock(p, &flags);
2239 old_state = p->state;
2240 if (!(old_state & state))
2248 this_cpu = smp_processor_id();
2251 if (unlikely(task_running(rq, p)))
2254 cpu = p->sched_class->select_task_rq(p, sync);
2255 if (cpu != orig_cpu) {
2256 set_task_cpu(p, cpu);
2257 task_rq_unlock(rq, &flags);
2258 /* might preempt at this point */
2259 rq = task_rq_lock(p, &flags);
2260 old_state = p->state;
2261 if (!(old_state & state))
2266 this_cpu = smp_processor_id();
2270 #ifdef CONFIG_SCHEDSTATS
2271 schedstat_inc(rq, ttwu_count);
2272 if (cpu == this_cpu)
2273 schedstat_inc(rq, ttwu_local);
2275 struct sched_domain *sd;
2276 for_each_domain(this_cpu, sd) {
2277 if (cpu_isset(cpu, sd->span)) {
2278 schedstat_inc(sd, ttwu_wake_remote);
2283 #endif /* CONFIG_SCHEDSTATS */
2286 #endif /* CONFIG_SMP */
2287 schedstat_inc(p, se.nr_wakeups);
2289 schedstat_inc(p, se.nr_wakeups_sync);
2290 if (orig_cpu != cpu)
2291 schedstat_inc(p, se.nr_wakeups_migrate);
2292 if (cpu == this_cpu)
2293 schedstat_inc(p, se.nr_wakeups_local);
2295 schedstat_inc(p, se.nr_wakeups_remote);
2296 update_rq_clock(rq);
2297 activate_task(rq, p, 1);
2301 trace_sched_wakeup(rq, p);
2302 check_preempt_curr(rq, p, sync);
2304 p->state = TASK_RUNNING;
2306 if (p->sched_class->task_wake_up)
2307 p->sched_class->task_wake_up(rq, p);
2310 current->se.last_wakeup = current->se.sum_exec_runtime;
2312 task_rq_unlock(rq, &flags);
2317 int wake_up_process(struct task_struct *p)
2319 return try_to_wake_up(p, TASK_ALL, 0);
2321 EXPORT_SYMBOL(wake_up_process);
2323 int wake_up_state(struct task_struct *p, unsigned int state)
2325 return try_to_wake_up(p, state, 0);
2329 * Perform scheduler related setup for a newly forked process p.
2330 * p is forked by current.
2332 * __sched_fork() is basic setup used by init_idle() too:
2334 static void __sched_fork(struct task_struct *p)
2336 p->se.exec_start = 0;
2337 p->se.sum_exec_runtime = 0;
2338 p->se.prev_sum_exec_runtime = 0;
2339 p->se.last_wakeup = 0;
2340 p->se.avg_overlap = 0;
2342 #ifdef CONFIG_SCHEDSTATS
2343 p->se.wait_start = 0;
2344 p->se.sum_sleep_runtime = 0;
2345 p->se.sleep_start = 0;
2346 p->se.block_start = 0;
2347 p->se.sleep_max = 0;
2348 p->se.block_max = 0;
2350 p->se.slice_max = 0;
2354 INIT_LIST_HEAD(&p->rt.run_list);
2356 INIT_LIST_HEAD(&p->se.group_node);
2358 #ifdef CONFIG_PREEMPT_NOTIFIERS
2359 INIT_HLIST_HEAD(&p->preempt_notifiers);
2363 * We mark the process as running here, but have not actually
2364 * inserted it onto the runqueue yet. This guarantees that
2365 * nobody will actually run it, and a signal or other external
2366 * event cannot wake it up and insert it on the runqueue either.
2368 p->state = TASK_RUNNING;
2372 * fork()/clone()-time setup:
2374 void sched_fork(struct task_struct *p, int clone_flags)
2376 int cpu = get_cpu();
2381 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2383 set_task_cpu(p, cpu);
2386 * Make sure we do not leak PI boosting priority to the child:
2388 p->prio = current->normal_prio;
2389 if (!rt_prio(p->prio))
2390 p->sched_class = &fair_sched_class;
2392 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2393 if (likely(sched_info_on()))
2394 memset(&p->sched_info, 0, sizeof(p->sched_info));
2396 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2399 #ifdef CONFIG_PREEMPT
2400 /* Want to start with kernel preemption disabled. */
2401 task_thread_info(p)->preempt_count = 1;
2407 * wake_up_new_task - wake up a newly created task for the first time.
2409 * This function will do some initial scheduler statistics housekeeping
2410 * that must be done for every newly created context, then puts the task
2411 * on the runqueue and wakes it.
2413 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2415 unsigned long flags;
2418 rq = task_rq_lock(p, &flags);
2419 BUG_ON(p->state != TASK_RUNNING);
2420 update_rq_clock(rq);
2422 p->prio = effective_prio(p);
2424 if (!p->sched_class->task_new || !current->se.on_rq) {
2425 activate_task(rq, p, 0);
2428 * Let the scheduling class do new task startup
2429 * management (if any):
2431 p->sched_class->task_new(rq, p);
2434 trace_sched_wakeup_new(rq, p);
2435 check_preempt_curr(rq, p, 0);
2437 if (p->sched_class->task_wake_up)
2438 p->sched_class->task_wake_up(rq, p);
2440 task_rq_unlock(rq, &flags);
2443 #ifdef CONFIG_PREEMPT_NOTIFIERS
2446 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2447 * @notifier: notifier struct to register
2449 void preempt_notifier_register(struct preempt_notifier *notifier)
2451 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2453 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2456 * preempt_notifier_unregister - no longer interested in preemption notifications
2457 * @notifier: notifier struct to unregister
2459 * This is safe to call from within a preemption notifier.
2461 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2463 hlist_del(¬ifier->link);
2465 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2467 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2469 struct preempt_notifier *notifier;
2470 struct hlist_node *node;
2472 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2473 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2477 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2478 struct task_struct *next)
2480 struct preempt_notifier *notifier;
2481 struct hlist_node *node;
2483 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2484 notifier->ops->sched_out(notifier, next);
2487 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2489 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2494 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2495 struct task_struct *next)
2499 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2502 * prepare_task_switch - prepare to switch tasks
2503 * @rq: the runqueue preparing to switch
2504 * @prev: the current task that is being switched out
2505 * @next: the task we are going to switch to.
2507 * This is called with the rq lock held and interrupts off. It must
2508 * be paired with a subsequent finish_task_switch after the context
2511 * prepare_task_switch sets up locking and calls architecture specific
2515 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2516 struct task_struct *next)
2518 fire_sched_out_preempt_notifiers(prev, next);
2519 prepare_lock_switch(rq, next);
2520 prepare_arch_switch(next);
2524 * finish_task_switch - clean up after a task-switch
2525 * @rq: runqueue associated with task-switch
2526 * @prev: the thread we just switched away from.
2528 * finish_task_switch must be called after the context switch, paired
2529 * with a prepare_task_switch call before the context switch.
2530 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2531 * and do any other architecture-specific cleanup actions.
2533 * Note that we may have delayed dropping an mm in context_switch(). If
2534 * so, we finish that here outside of the runqueue lock. (Doing it
2535 * with the lock held can cause deadlocks; see schedule() for
2538 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2539 __releases(rq->lock)
2541 struct mm_struct *mm = rq->prev_mm;
2547 * A task struct has one reference for the use as "current".
2548 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2549 * schedule one last time. The schedule call will never return, and
2550 * the scheduled task must drop that reference.
2551 * The test for TASK_DEAD must occur while the runqueue locks are
2552 * still held, otherwise prev could be scheduled on another cpu, die
2553 * there before we look at prev->state, and then the reference would
2555 * Manfred Spraul <manfred@colorfullife.com>
2557 prev_state = prev->state;
2558 finish_arch_switch(prev);
2559 finish_lock_switch(rq, prev);
2561 if (current->sched_class->post_schedule)
2562 current->sched_class->post_schedule(rq);
2565 fire_sched_in_preempt_notifiers(current);
2568 if (unlikely(prev_state == TASK_DEAD)) {
2570 * Remove function-return probe instances associated with this
2571 * task and put them back on the free list.
2573 kprobe_flush_task(prev);
2574 put_task_struct(prev);
2579 * schedule_tail - first thing a freshly forked thread must call.
2580 * @prev: the thread we just switched away from.
2582 asmlinkage void schedule_tail(struct task_struct *prev)
2583 __releases(rq->lock)
2585 struct rq *rq = this_rq();
2587 finish_task_switch(rq, prev);
2588 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2589 /* In this case, finish_task_switch does not reenable preemption */
2592 if (current->set_child_tid)
2593 put_user(task_pid_vnr(current), current->set_child_tid);
2597 * context_switch - switch to the new MM and the new
2598 * thread's register state.
2601 context_switch(struct rq *rq, struct task_struct *prev,
2602 struct task_struct *next)
2604 struct mm_struct *mm, *oldmm;
2606 prepare_task_switch(rq, prev, next);
2607 trace_sched_switch(rq, prev, next);
2609 oldmm = prev->active_mm;
2611 * For paravirt, this is coupled with an exit in switch_to to
2612 * combine the page table reload and the switch backend into
2615 arch_enter_lazy_cpu_mode();
2617 if (unlikely(!mm)) {
2618 next->active_mm = oldmm;
2619 atomic_inc(&oldmm->mm_count);
2620 enter_lazy_tlb(oldmm, next);
2622 switch_mm(oldmm, mm, next);
2624 if (unlikely(!prev->mm)) {
2625 prev->active_mm = NULL;
2626 rq->prev_mm = oldmm;
2629 * Since the runqueue lock will be released by the next
2630 * task (which is an invalid locking op but in the case
2631 * of the scheduler it's an obvious special-case), so we
2632 * do an early lockdep release here:
2634 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2635 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2638 /* Here we just switch the register state and the stack. */
2639 switch_to(prev, next, prev);
2643 * this_rq must be evaluated again because prev may have moved
2644 * CPUs since it called schedule(), thus the 'rq' on its stack
2645 * frame will be invalid.
2647 finish_task_switch(this_rq(), prev);
2651 * nr_running, nr_uninterruptible and nr_context_switches:
2653 * externally visible scheduler statistics: current number of runnable
2654 * threads, current number of uninterruptible-sleeping threads, total
2655 * number of context switches performed since bootup.
2657 unsigned long nr_running(void)
2659 unsigned long i, sum = 0;
2661 for_each_online_cpu(i)
2662 sum += cpu_rq(i)->nr_running;
2667 unsigned long nr_uninterruptible(void)
2669 unsigned long i, sum = 0;
2671 for_each_possible_cpu(i)
2672 sum += cpu_rq(i)->nr_uninterruptible;
2675 * Since we read the counters lockless, it might be slightly
2676 * inaccurate. Do not allow it to go below zero though:
2678 if (unlikely((long)sum < 0))
2684 unsigned long long nr_context_switches(void)
2687 unsigned long long sum = 0;
2689 for_each_possible_cpu(i)
2690 sum += cpu_rq(i)->nr_switches;
2695 unsigned long nr_iowait(void)
2697 unsigned long i, sum = 0;
2699 for_each_possible_cpu(i)
2700 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2705 unsigned long nr_active(void)
2707 unsigned long i, running = 0, uninterruptible = 0;
2709 for_each_online_cpu(i) {
2710 running += cpu_rq(i)->nr_running;
2711 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2714 if (unlikely((long)uninterruptible < 0))
2715 uninterruptible = 0;
2717 return running + uninterruptible;
2721 * Update rq->cpu_load[] statistics. This function is usually called every
2722 * scheduler tick (TICK_NSEC).
2724 static void update_cpu_load(struct rq *this_rq)
2726 unsigned long this_load = this_rq->load.weight;
2729 this_rq->nr_load_updates++;
2731 /* Update our load: */
2732 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2733 unsigned long old_load, new_load;
2735 /* scale is effectively 1 << i now, and >> i divides by scale */
2737 old_load = this_rq->cpu_load[i];
2738 new_load = this_load;
2740 * Round up the averaging division if load is increasing. This
2741 * prevents us from getting stuck on 9 if the load is 10, for
2744 if (new_load > old_load)
2745 new_load += scale-1;
2746 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2753 * double_rq_lock - safely lock two runqueues
2755 * Note this does not disable interrupts like task_rq_lock,
2756 * you need to do so manually before calling.
2758 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2759 __acquires(rq1->lock)
2760 __acquires(rq2->lock)
2762 BUG_ON(!irqs_disabled());
2764 spin_lock(&rq1->lock);
2765 __acquire(rq2->lock); /* Fake it out ;) */
2768 spin_lock(&rq1->lock);
2769 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2771 spin_lock(&rq2->lock);
2772 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2775 update_rq_clock(rq1);
2776 update_rq_clock(rq2);
2780 * double_rq_unlock - safely unlock two runqueues
2782 * Note this does not restore interrupts like task_rq_unlock,
2783 * you need to do so manually after calling.
2785 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2786 __releases(rq1->lock)
2787 __releases(rq2->lock)
2789 spin_unlock(&rq1->lock);
2791 spin_unlock(&rq2->lock);
2793 __release(rq2->lock);
2797 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2799 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2800 __releases(this_rq->lock)
2801 __acquires(busiest->lock)
2802 __acquires(this_rq->lock)
2806 if (unlikely(!irqs_disabled())) {
2807 /* printk() doesn't work good under rq->lock */
2808 spin_unlock(&this_rq->lock);
2811 if (unlikely(!spin_trylock(&busiest->lock))) {
2812 if (busiest < this_rq) {
2813 spin_unlock(&this_rq->lock);
2814 spin_lock(&busiest->lock);
2815 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
2818 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
2823 static void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2824 __releases(busiest->lock)
2826 spin_unlock(&busiest->lock);
2827 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
2831 * If dest_cpu is allowed for this process, migrate the task to it.
2832 * This is accomplished by forcing the cpu_allowed mask to only
2833 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2834 * the cpu_allowed mask is restored.
2836 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2838 struct migration_req req;
2839 unsigned long flags;
2842 rq = task_rq_lock(p, &flags);
2843 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2844 || unlikely(!cpu_active(dest_cpu)))
2847 trace_sched_migrate_task(rq, p, dest_cpu);
2848 /* force the process onto the specified CPU */
2849 if (migrate_task(p, dest_cpu, &req)) {
2850 /* Need to wait for migration thread (might exit: take ref). */
2851 struct task_struct *mt = rq->migration_thread;
2853 get_task_struct(mt);
2854 task_rq_unlock(rq, &flags);
2855 wake_up_process(mt);
2856 put_task_struct(mt);
2857 wait_for_completion(&req.done);
2862 task_rq_unlock(rq, &flags);
2866 * sched_exec - execve() is a valuable balancing opportunity, because at
2867 * this point the task has the smallest effective memory and cache footprint.
2869 void sched_exec(void)
2871 int new_cpu, this_cpu = get_cpu();
2872 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2874 if (new_cpu != this_cpu)
2875 sched_migrate_task(current, new_cpu);
2879 * pull_task - move a task from a remote runqueue to the local runqueue.
2880 * Both runqueues must be locked.
2882 static void pull_task(struct rq *src_rq, struct task_struct *p,
2883 struct rq *this_rq, int this_cpu)
2885 deactivate_task(src_rq, p, 0);
2886 set_task_cpu(p, this_cpu);
2887 activate_task(this_rq, p, 0);
2889 * Note that idle threads have a prio of MAX_PRIO, for this test
2890 * to be always true for them.
2892 check_preempt_curr(this_rq, p, 0);
2896 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2899 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2900 struct sched_domain *sd, enum cpu_idle_type idle,
2904 * We do not migrate tasks that are:
2905 * 1) running (obviously), or
2906 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2907 * 3) are cache-hot on their current CPU.
2909 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2910 schedstat_inc(p, se.nr_failed_migrations_affine);
2915 if (task_running(rq, p)) {
2916 schedstat_inc(p, se.nr_failed_migrations_running);
2921 * Aggressive migration if:
2922 * 1) task is cache cold, or
2923 * 2) too many balance attempts have failed.
2926 if (!task_hot(p, rq->clock, sd) ||
2927 sd->nr_balance_failed > sd->cache_nice_tries) {
2928 #ifdef CONFIG_SCHEDSTATS
2929 if (task_hot(p, rq->clock, sd)) {
2930 schedstat_inc(sd, lb_hot_gained[idle]);
2931 schedstat_inc(p, se.nr_forced_migrations);
2937 if (task_hot(p, rq->clock, sd)) {
2938 schedstat_inc(p, se.nr_failed_migrations_hot);
2944 static unsigned long
2945 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2946 unsigned long max_load_move, struct sched_domain *sd,
2947 enum cpu_idle_type idle, int *all_pinned,
2948 int *this_best_prio, struct rq_iterator *iterator)
2950 int loops = 0, pulled = 0, pinned = 0;
2951 struct task_struct *p;
2952 long rem_load_move = max_load_move;
2954 if (max_load_move == 0)
2960 * Start the load-balancing iterator:
2962 p = iterator->start(iterator->arg);
2964 if (!p || loops++ > sysctl_sched_nr_migrate)
2967 if ((p->se.load.weight >> 1) > rem_load_move ||
2968 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2969 p = iterator->next(iterator->arg);
2973 pull_task(busiest, p, this_rq, this_cpu);
2975 rem_load_move -= p->se.load.weight;
2978 * We only want to steal up to the prescribed amount of weighted load.
2980 if (rem_load_move > 0) {
2981 if (p->prio < *this_best_prio)
2982 *this_best_prio = p->prio;
2983 p = iterator->next(iterator->arg);
2988 * Right now, this is one of only two places pull_task() is called,
2989 * so we can safely collect pull_task() stats here rather than
2990 * inside pull_task().
2992 schedstat_add(sd, lb_gained[idle], pulled);
2995 *all_pinned = pinned;
2997 return max_load_move - rem_load_move;
3001 * move_tasks tries to move up to max_load_move weighted load from busiest to
3002 * this_rq, as part of a balancing operation within domain "sd".
3003 * Returns 1 if successful and 0 otherwise.
3005 * Called with both runqueues locked.
3007 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3008 unsigned long max_load_move,
3009 struct sched_domain *sd, enum cpu_idle_type idle,
3012 const struct sched_class *class = sched_class_highest;
3013 unsigned long total_load_moved = 0;
3014 int this_best_prio = this_rq->curr->prio;
3018 class->load_balance(this_rq, this_cpu, busiest,
3019 max_load_move - total_load_moved,
3020 sd, idle, all_pinned, &this_best_prio);
3021 class = class->next;
3023 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3026 } while (class && max_load_move > total_load_moved);
3028 return total_load_moved > 0;
3032 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3033 struct sched_domain *sd, enum cpu_idle_type idle,
3034 struct rq_iterator *iterator)
3036 struct task_struct *p = iterator->start(iterator->arg);
3040 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3041 pull_task(busiest, p, this_rq, this_cpu);
3043 * Right now, this is only the second place pull_task()
3044 * is called, so we can safely collect pull_task()
3045 * stats here rather than inside pull_task().
3047 schedstat_inc(sd, lb_gained[idle]);
3051 p = iterator->next(iterator->arg);
3058 * move_one_task tries to move exactly one task from busiest to this_rq, as
3059 * part of active balancing operations within "domain".
3060 * Returns 1 if successful and 0 otherwise.
3062 * Called with both runqueues locked.
3064 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3065 struct sched_domain *sd, enum cpu_idle_type idle)
3067 const struct sched_class *class;
3069 for (class = sched_class_highest; class; class = class->next)
3070 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3077 * find_busiest_group finds and returns the busiest CPU group within the
3078 * domain. It calculates and returns the amount of weighted load which
3079 * should be moved to restore balance via the imbalance parameter.
3081 static struct sched_group *
3082 find_busiest_group(struct sched_domain *sd, int this_cpu,
3083 unsigned long *imbalance, enum cpu_idle_type idle,
3084 int *sd_idle, const cpumask_t *cpus, int *balance)
3086 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3087 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3088 unsigned long max_pull;
3089 unsigned long busiest_load_per_task, busiest_nr_running;
3090 unsigned long this_load_per_task, this_nr_running;
3091 int load_idx, group_imb = 0;
3092 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3093 int power_savings_balance = 1;
3094 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3095 unsigned long min_nr_running = ULONG_MAX;
3096 struct sched_group *group_min = NULL, *group_leader = NULL;
3099 max_load = this_load = total_load = total_pwr = 0;
3100 busiest_load_per_task = busiest_nr_running = 0;
3101 this_load_per_task = this_nr_running = 0;
3103 if (idle == CPU_NOT_IDLE)
3104 load_idx = sd->busy_idx;
3105 else if (idle == CPU_NEWLY_IDLE)
3106 load_idx = sd->newidle_idx;
3108 load_idx = sd->idle_idx;
3111 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3114 int __group_imb = 0;
3115 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3116 unsigned long sum_nr_running, sum_weighted_load;
3117 unsigned long sum_avg_load_per_task;
3118 unsigned long avg_load_per_task;
3120 local_group = cpu_isset(this_cpu, group->cpumask);
3123 balance_cpu = first_cpu(group->cpumask);
3125 /* Tally up the load of all CPUs in the group */
3126 sum_weighted_load = sum_nr_running = avg_load = 0;
3127 sum_avg_load_per_task = avg_load_per_task = 0;
3130 min_cpu_load = ~0UL;
3132 for_each_cpu_mask_nr(i, group->cpumask) {
3135 if (!cpu_isset(i, *cpus))
3140 if (*sd_idle && rq->nr_running)
3143 /* Bias balancing toward cpus of our domain */
3145 if (idle_cpu(i) && !first_idle_cpu) {
3150 load = target_load(i, load_idx);
3152 load = source_load(i, load_idx);
3153 if (load > max_cpu_load)
3154 max_cpu_load = load;
3155 if (min_cpu_load > load)
3156 min_cpu_load = load;
3160 sum_nr_running += rq->nr_running;
3161 sum_weighted_load += weighted_cpuload(i);
3163 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3167 * First idle cpu or the first cpu(busiest) in this sched group
3168 * is eligible for doing load balancing at this and above
3169 * domains. In the newly idle case, we will allow all the cpu's
3170 * to do the newly idle load balance.
3172 if (idle != CPU_NEWLY_IDLE && local_group &&
3173 balance_cpu != this_cpu && balance) {
3178 total_load += avg_load;
3179 total_pwr += group->__cpu_power;
3181 /* Adjust by relative CPU power of the group */
3182 avg_load = sg_div_cpu_power(group,
3183 avg_load * SCHED_LOAD_SCALE);
3187 * Consider the group unbalanced when the imbalance is larger
3188 * than the average weight of two tasks.
3190 * APZ: with cgroup the avg task weight can vary wildly and
3191 * might not be a suitable number - should we keep a
3192 * normalized nr_running number somewhere that negates
3195 avg_load_per_task = sg_div_cpu_power(group,
3196 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3198 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3201 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3204 this_load = avg_load;
3206 this_nr_running = sum_nr_running;
3207 this_load_per_task = sum_weighted_load;
3208 } else if (avg_load > max_load &&
3209 (sum_nr_running > group_capacity || __group_imb)) {
3210 max_load = avg_load;
3212 busiest_nr_running = sum_nr_running;
3213 busiest_load_per_task = sum_weighted_load;
3214 group_imb = __group_imb;
3217 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3219 * Busy processors will not participate in power savings
3222 if (idle == CPU_NOT_IDLE ||
3223 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3227 * If the local group is idle or completely loaded
3228 * no need to do power savings balance at this domain
3230 if (local_group && (this_nr_running >= group_capacity ||
3232 power_savings_balance = 0;
3235 * If a group is already running at full capacity or idle,
3236 * don't include that group in power savings calculations
3238 if (!power_savings_balance || sum_nr_running >= group_capacity
3243 * Calculate the group which has the least non-idle load.
3244 * This is the group from where we need to pick up the load
3247 if ((sum_nr_running < min_nr_running) ||
3248 (sum_nr_running == min_nr_running &&
3249 first_cpu(group->cpumask) <
3250 first_cpu(group_min->cpumask))) {
3252 min_nr_running = sum_nr_running;
3253 min_load_per_task = sum_weighted_load /
3258 * Calculate the group which is almost near its
3259 * capacity but still has some space to pick up some load
3260 * from other group and save more power
3262 if (sum_nr_running <= group_capacity - 1) {
3263 if (sum_nr_running > leader_nr_running ||
3264 (sum_nr_running == leader_nr_running &&
3265 first_cpu(group->cpumask) >
3266 first_cpu(group_leader->cpumask))) {
3267 group_leader = group;
3268 leader_nr_running = sum_nr_running;
3273 group = group->next;
3274 } while (group != sd->groups);
3276 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3279 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3281 if (this_load >= avg_load ||
3282 100*max_load <= sd->imbalance_pct*this_load)
3285 busiest_load_per_task /= busiest_nr_running;
3287 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3290 * We're trying to get all the cpus to the average_load, so we don't
3291 * want to push ourselves above the average load, nor do we wish to
3292 * reduce the max loaded cpu below the average load, as either of these
3293 * actions would just result in more rebalancing later, and ping-pong
3294 * tasks around. Thus we look for the minimum possible imbalance.
3295 * Negative imbalances (*we* are more loaded than anyone else) will
3296 * be counted as no imbalance for these purposes -- we can't fix that
3297 * by pulling tasks to us. Be careful of negative numbers as they'll
3298 * appear as very large values with unsigned longs.
3300 if (max_load <= busiest_load_per_task)
3304 * In the presence of smp nice balancing, certain scenarios can have
3305 * max load less than avg load(as we skip the groups at or below
3306 * its cpu_power, while calculating max_load..)
3308 if (max_load < avg_load) {
3310 goto small_imbalance;
3313 /* Don't want to pull so many tasks that a group would go idle */
3314 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3316 /* How much load to actually move to equalise the imbalance */
3317 *imbalance = min(max_pull * busiest->__cpu_power,
3318 (avg_load - this_load) * this->__cpu_power)
3322 * if *imbalance is less than the average load per runnable task
3323 * there is no gaurantee that any tasks will be moved so we'll have
3324 * a think about bumping its value to force at least one task to be
3327 if (*imbalance < busiest_load_per_task) {
3328 unsigned long tmp, pwr_now, pwr_move;
3332 pwr_move = pwr_now = 0;
3334 if (this_nr_running) {
3335 this_load_per_task /= this_nr_running;
3336 if (busiest_load_per_task > this_load_per_task)
3339 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3341 if (max_load - this_load + 2*busiest_load_per_task >=
3342 busiest_load_per_task * imbn) {
3343 *imbalance = busiest_load_per_task;
3348 * OK, we don't have enough imbalance to justify moving tasks,
3349 * however we may be able to increase total CPU power used by
3353 pwr_now += busiest->__cpu_power *
3354 min(busiest_load_per_task, max_load);
3355 pwr_now += this->__cpu_power *
3356 min(this_load_per_task, this_load);
3357 pwr_now /= SCHED_LOAD_SCALE;
3359 /* Amount of load we'd subtract */
3360 tmp = sg_div_cpu_power(busiest,
3361 busiest_load_per_task * SCHED_LOAD_SCALE);
3363 pwr_move += busiest->__cpu_power *
3364 min(busiest_load_per_task, max_load - tmp);
3366 /* Amount of load we'd add */
3367 if (max_load * busiest->__cpu_power <
3368 busiest_load_per_task * SCHED_LOAD_SCALE)
3369 tmp = sg_div_cpu_power(this,
3370 max_load * busiest->__cpu_power);
3372 tmp = sg_div_cpu_power(this,
3373 busiest_load_per_task * SCHED_LOAD_SCALE);
3374 pwr_move += this->__cpu_power *
3375 min(this_load_per_task, this_load + tmp);
3376 pwr_move /= SCHED_LOAD_SCALE;
3378 /* Move if we gain throughput */
3379 if (pwr_move > pwr_now)
3380 *imbalance = busiest_load_per_task;
3386 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3387 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3390 if (this == group_leader && group_leader != group_min) {
3391 *imbalance = min_load_per_task;
3401 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3404 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3405 unsigned long imbalance, const cpumask_t *cpus)
3407 struct rq *busiest = NULL, *rq;
3408 unsigned long max_load = 0;
3411 for_each_cpu_mask_nr(i, group->cpumask) {
3414 if (!cpu_isset(i, *cpus))
3418 wl = weighted_cpuload(i);
3420 if (rq->nr_running == 1 && wl > imbalance)
3423 if (wl > max_load) {
3433 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3434 * so long as it is large enough.
3436 #define MAX_PINNED_INTERVAL 512
3439 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3440 * tasks if there is an imbalance.
3442 static int load_balance(int this_cpu, struct rq *this_rq,
3443 struct sched_domain *sd, enum cpu_idle_type idle,
3444 int *balance, cpumask_t *cpus)
3446 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3447 struct sched_group *group;
3448 unsigned long imbalance;
3450 unsigned long flags;
3455 * When power savings policy is enabled for the parent domain, idle
3456 * sibling can pick up load irrespective of busy siblings. In this case,
3457 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3458 * portraying it as CPU_NOT_IDLE.
3460 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3461 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3464 schedstat_inc(sd, lb_count[idle]);
3468 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3475 schedstat_inc(sd, lb_nobusyg[idle]);
3479 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3481 schedstat_inc(sd, lb_nobusyq[idle]);
3485 BUG_ON(busiest == this_rq);
3487 schedstat_add(sd, lb_imbalance[idle], imbalance);
3490 if (busiest->nr_running > 1) {
3492 * Attempt to move tasks. If find_busiest_group has found
3493 * an imbalance but busiest->nr_running <= 1, the group is
3494 * still unbalanced. ld_moved simply stays zero, so it is
3495 * correctly treated as an imbalance.
3497 local_irq_save(flags);
3498 double_rq_lock(this_rq, busiest);
3499 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3500 imbalance, sd, idle, &all_pinned);
3501 double_rq_unlock(this_rq, busiest);
3502 local_irq_restore(flags);
3505 * some other cpu did the load balance for us.
3507 if (ld_moved && this_cpu != smp_processor_id())
3508 resched_cpu(this_cpu);
3510 /* All tasks on this runqueue were pinned by CPU affinity */
3511 if (unlikely(all_pinned)) {
3512 cpu_clear(cpu_of(busiest), *cpus);
3513 if (!cpus_empty(*cpus))
3520 schedstat_inc(sd, lb_failed[idle]);
3521 sd->nr_balance_failed++;
3523 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3525 spin_lock_irqsave(&busiest->lock, flags);
3527 /* don't kick the migration_thread, if the curr
3528 * task on busiest cpu can't be moved to this_cpu
3530 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3531 spin_unlock_irqrestore(&busiest->lock, flags);
3533 goto out_one_pinned;
3536 if (!busiest->active_balance) {
3537 busiest->active_balance = 1;
3538 busiest->push_cpu = this_cpu;
3541 spin_unlock_irqrestore(&busiest->lock, flags);
3543 wake_up_process(busiest->migration_thread);
3546 * We've kicked active balancing, reset the failure
3549 sd->nr_balance_failed = sd->cache_nice_tries+1;
3552 sd->nr_balance_failed = 0;
3554 if (likely(!active_balance)) {
3555 /* We were unbalanced, so reset the balancing interval */
3556 sd->balance_interval = sd->min_interval;
3559 * If we've begun active balancing, start to back off. This
3560 * case may not be covered by the all_pinned logic if there
3561 * is only 1 task on the busy runqueue (because we don't call
3564 if (sd->balance_interval < sd->max_interval)
3565 sd->balance_interval *= 2;
3568 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3569 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3575 schedstat_inc(sd, lb_balanced[idle]);
3577 sd->nr_balance_failed = 0;
3580 /* tune up the balancing interval */
3581 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3582 (sd->balance_interval < sd->max_interval))
3583 sd->balance_interval *= 2;
3585 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3586 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3597 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3598 * tasks if there is an imbalance.
3600 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3601 * this_rq is locked.
3604 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3607 struct sched_group *group;
3608 struct rq *busiest = NULL;
3609 unsigned long imbalance;
3617 * When power savings policy is enabled for the parent domain, idle
3618 * sibling can pick up load irrespective of busy siblings. In this case,
3619 * let the state of idle sibling percolate up as IDLE, instead of
3620 * portraying it as CPU_NOT_IDLE.
3622 if (sd->flags & SD_SHARE_CPUPOWER &&
3623 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3626 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3628 update_shares_locked(this_rq, sd);
3629 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3630 &sd_idle, cpus, NULL);
3632 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3636 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3638 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3642 BUG_ON(busiest == this_rq);
3644 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3647 if (busiest->nr_running > 1) {
3648 /* Attempt to move tasks */
3649 double_lock_balance(this_rq, busiest);
3650 /* this_rq->clock is already updated */
3651 update_rq_clock(busiest);
3652 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3653 imbalance, sd, CPU_NEWLY_IDLE,
3655 double_unlock_balance(this_rq, busiest);
3657 if (unlikely(all_pinned)) {
3658 cpu_clear(cpu_of(busiest), *cpus);
3659 if (!cpus_empty(*cpus))
3665 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3666 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3667 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3670 sd->nr_balance_failed = 0;
3672 update_shares_locked(this_rq, sd);
3676 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3677 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3678 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3680 sd->nr_balance_failed = 0;
3686 * idle_balance is called by schedule() if this_cpu is about to become
3687 * idle. Attempts to pull tasks from other CPUs.
3689 static void idle_balance(int this_cpu, struct rq *this_rq)
3691 struct sched_domain *sd;
3692 int pulled_task = -1;
3693 unsigned long next_balance = jiffies + HZ;
3696 for_each_domain(this_cpu, sd) {
3697 unsigned long interval;
3699 if (!(sd->flags & SD_LOAD_BALANCE))
3702 if (sd->flags & SD_BALANCE_NEWIDLE)
3703 /* If we've pulled tasks over stop searching: */
3704 pulled_task = load_balance_newidle(this_cpu, this_rq,
3707 interval = msecs_to_jiffies(sd->balance_interval);
3708 if (time_after(next_balance, sd->last_balance + interval))
3709 next_balance = sd->last_balance + interval;
3713 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3715 * We are going idle. next_balance may be set based on
3716 * a busy processor. So reset next_balance.
3718 this_rq->next_balance = next_balance;
3723 * active_load_balance is run by migration threads. It pushes running tasks
3724 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3725 * running on each physical CPU where possible, and avoids physical /
3726 * logical imbalances.
3728 * Called with busiest_rq locked.
3730 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3732 int target_cpu = busiest_rq->push_cpu;
3733 struct sched_domain *sd;
3734 struct rq *target_rq;
3736 /* Is there any task to move? */
3737 if (busiest_rq->nr_running <= 1)
3740 target_rq = cpu_rq(target_cpu);
3743 * This condition is "impossible", if it occurs
3744 * we need to fix it. Originally reported by
3745 * Bjorn Helgaas on a 128-cpu setup.
3747 BUG_ON(busiest_rq == target_rq);
3749 /* move a task from busiest_rq to target_rq */
3750 double_lock_balance(busiest_rq, target_rq);
3751 update_rq_clock(busiest_rq);
3752 update_rq_clock(target_rq);
3754 /* Search for an sd spanning us and the target CPU. */
3755 for_each_domain(target_cpu, sd) {
3756 if ((sd->flags & SD_LOAD_BALANCE) &&
3757 cpu_isset(busiest_cpu, sd->span))
3762 schedstat_inc(sd, alb_count);
3764 if (move_one_task(target_rq, target_cpu, busiest_rq,
3766 schedstat_inc(sd, alb_pushed);
3768 schedstat_inc(sd, alb_failed);
3770 double_unlock_balance(busiest_rq, target_rq);
3775 atomic_t load_balancer;
3777 } nohz ____cacheline_aligned = {
3778 .load_balancer = ATOMIC_INIT(-1),
3779 .cpu_mask = CPU_MASK_NONE,
3783 * This routine will try to nominate the ilb (idle load balancing)
3784 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3785 * load balancing on behalf of all those cpus. If all the cpus in the system
3786 * go into this tickless mode, then there will be no ilb owner (as there is
3787 * no need for one) and all the cpus will sleep till the next wakeup event
3790 * For the ilb owner, tick is not stopped. And this tick will be used
3791 * for idle load balancing. ilb owner will still be part of
3794 * While stopping the tick, this cpu will become the ilb owner if there
3795 * is no other owner. And will be the owner till that cpu becomes busy
3796 * or if all cpus in the system stop their ticks at which point
3797 * there is no need for ilb owner.
3799 * When the ilb owner becomes busy, it nominates another owner, during the
3800 * next busy scheduler_tick()
3802 int select_nohz_load_balancer(int stop_tick)
3804 int cpu = smp_processor_id();
3807 cpu_set(cpu, nohz.cpu_mask);
3808 cpu_rq(cpu)->in_nohz_recently = 1;
3811 * If we are going offline and still the leader, give up!
3813 if (!cpu_active(cpu) &&
3814 atomic_read(&nohz.load_balancer) == cpu) {
3815 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3820 /* time for ilb owner also to sleep */
3821 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3822 if (atomic_read(&nohz.load_balancer) == cpu)
3823 atomic_set(&nohz.load_balancer, -1);
3827 if (atomic_read(&nohz.load_balancer) == -1) {
3828 /* make me the ilb owner */
3829 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3831 } else if (atomic_read(&nohz.load_balancer) == cpu)
3834 if (!cpu_isset(cpu, nohz.cpu_mask))
3837 cpu_clear(cpu, nohz.cpu_mask);
3839 if (atomic_read(&nohz.load_balancer) == cpu)
3840 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3847 static DEFINE_SPINLOCK(balancing);
3850 * It checks each scheduling domain to see if it is due to be balanced,
3851 * and initiates a balancing operation if so.
3853 * Balancing parameters are set up in arch_init_sched_domains.
3855 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3858 struct rq *rq = cpu_rq(cpu);
3859 unsigned long interval;
3860 struct sched_domain *sd;
3861 /* Earliest time when we have to do rebalance again */
3862 unsigned long next_balance = jiffies + 60*HZ;
3863 int update_next_balance = 0;
3867 for_each_domain(cpu, sd) {
3868 if (!(sd->flags & SD_LOAD_BALANCE))
3871 interval = sd->balance_interval;
3872 if (idle != CPU_IDLE)
3873 interval *= sd->busy_factor;
3875 /* scale ms to jiffies */
3876 interval = msecs_to_jiffies(interval);
3877 if (unlikely(!interval))
3879 if (interval > HZ*NR_CPUS/10)
3880 interval = HZ*NR_CPUS/10;
3882 need_serialize = sd->flags & SD_SERIALIZE;
3884 if (need_serialize) {
3885 if (!spin_trylock(&balancing))
3889 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3890 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3892 * We've pulled tasks over so either we're no
3893 * longer idle, or one of our SMT siblings is
3896 idle = CPU_NOT_IDLE;
3898 sd->last_balance = jiffies;
3901 spin_unlock(&balancing);
3903 if (time_after(next_balance, sd->last_balance + interval)) {
3904 next_balance = sd->last_balance + interval;
3905 update_next_balance = 1;
3909 * Stop the load balance at this level. There is another
3910 * CPU in our sched group which is doing load balancing more
3918 * next_balance will be updated only when there is a need.
3919 * When the cpu is attached to null domain for ex, it will not be
3922 if (likely(update_next_balance))
3923 rq->next_balance = next_balance;
3927 * run_rebalance_domains is triggered when needed from the scheduler tick.
3928 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3929 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3931 static void run_rebalance_domains(struct softirq_action *h)
3933 int this_cpu = smp_processor_id();
3934 struct rq *this_rq = cpu_rq(this_cpu);
3935 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3936 CPU_IDLE : CPU_NOT_IDLE;
3938 rebalance_domains(this_cpu, idle);
3942 * If this cpu is the owner for idle load balancing, then do the
3943 * balancing on behalf of the other idle cpus whose ticks are
3946 if (this_rq->idle_at_tick &&
3947 atomic_read(&nohz.load_balancer) == this_cpu) {
3948 cpumask_t cpus = nohz.cpu_mask;
3952 cpu_clear(this_cpu, cpus);
3953 for_each_cpu_mask_nr(balance_cpu, cpus) {
3955 * If this cpu gets work to do, stop the load balancing
3956 * work being done for other cpus. Next load
3957 * balancing owner will pick it up.
3962 rebalance_domains(balance_cpu, CPU_IDLE);
3964 rq = cpu_rq(balance_cpu);
3965 if (time_after(this_rq->next_balance, rq->next_balance))
3966 this_rq->next_balance = rq->next_balance;
3973 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3975 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3976 * idle load balancing owner or decide to stop the periodic load balancing,
3977 * if the whole system is idle.
3979 static inline void trigger_load_balance(struct rq *rq, int cpu)
3983 * If we were in the nohz mode recently and busy at the current
3984 * scheduler tick, then check if we need to nominate new idle
3987 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3988 rq->in_nohz_recently = 0;
3990 if (atomic_read(&nohz.load_balancer) == cpu) {
3991 cpu_clear(cpu, nohz.cpu_mask);
3992 atomic_set(&nohz.load_balancer, -1);
3995 if (atomic_read(&nohz.load_balancer) == -1) {
3997 * simple selection for now: Nominate the
3998 * first cpu in the nohz list to be the next
4001 * TBD: Traverse the sched domains and nominate
4002 * the nearest cpu in the nohz.cpu_mask.
4004 int ilb = first_cpu(nohz.cpu_mask);
4006 if (ilb < nr_cpu_ids)
4012 * If this cpu is idle and doing idle load balancing for all the
4013 * cpus with ticks stopped, is it time for that to stop?
4015 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4016 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4022 * If this cpu is idle and the idle load balancing is done by
4023 * someone else, then no need raise the SCHED_SOFTIRQ
4025 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4026 cpu_isset(cpu, nohz.cpu_mask))
4029 if (time_after_eq(jiffies, rq->next_balance))
4030 raise_softirq(SCHED_SOFTIRQ);
4033 #else /* CONFIG_SMP */
4036 * on UP we do not need to balance between CPUs:
4038 static inline void idle_balance(int cpu, struct rq *rq)
4044 DEFINE_PER_CPU(struct kernel_stat, kstat);
4046 EXPORT_PER_CPU_SYMBOL(kstat);
4049 * Return any ns on the sched_clock that have not yet been banked in
4050 * @p in case that task is currently running.
4052 unsigned long long task_delta_exec(struct task_struct *p)
4054 unsigned long flags;
4058 rq = task_rq_lock(p, &flags);
4060 if (task_current(rq, p)) {
4063 update_rq_clock(rq);
4064 delta_exec = rq->clock - p->se.exec_start;
4065 if ((s64)delta_exec > 0)
4069 task_rq_unlock(rq, &flags);
4075 * Account user cpu time to a process.
4076 * @p: the process that the cpu time gets accounted to
4077 * @cputime: the cpu time spent in user space since the last update
4079 void account_user_time(struct task_struct *p, cputime_t cputime)
4081 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4084 p->utime = cputime_add(p->utime, cputime);
4085 account_group_user_time(p, cputime);
4087 /* Add user time to cpustat. */
4088 tmp = cputime_to_cputime64(cputime);
4089 if (TASK_NICE(p) > 0)
4090 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4092 cpustat->user = cputime64_add(cpustat->user, tmp);
4093 /* Account for user time used */
4094 acct_update_integrals(p);
4098 * Account guest cpu time to a process.
4099 * @p: the process that the cpu time gets accounted to
4100 * @cputime: the cpu time spent in virtual machine since the last update
4102 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4105 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4107 tmp = cputime_to_cputime64(cputime);
4109 p->utime = cputime_add(p->utime, cputime);
4110 account_group_user_time(p, cputime);
4111 p->gtime = cputime_add(p->gtime, cputime);
4113 cpustat->user = cputime64_add(cpustat->user, tmp);
4114 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4118 * Account scaled user cpu time to a process.
4119 * @p: the process that the cpu time gets accounted to
4120 * @cputime: the cpu time spent in user space since the last update
4122 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4124 p->utimescaled = cputime_add(p->utimescaled, cputime);
4128 * Account system cpu time to a process.
4129 * @p: the process that the cpu time gets accounted to
4130 * @hardirq_offset: the offset to subtract from hardirq_count()
4131 * @cputime: the cpu time spent in kernel space since the last update
4133 void account_system_time(struct task_struct *p, int hardirq_offset,
4136 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4137 struct rq *rq = this_rq();
4140 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4141 account_guest_time(p, cputime);
4145 p->stime = cputime_add(p->stime, cputime);
4146 account_group_system_time(p, 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 account_group_system_time(p, steal);
4189 if (atomic_read(&rq->nr_iowait) > 0)
4190 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4192 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4194 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4198 * Use precise platform statistics if available:
4200 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4201 cputime_t task_utime(struct task_struct *p)
4206 cputime_t task_stime(struct task_struct *p)
4211 cputime_t task_utime(struct task_struct *p)
4213 clock_t utime = cputime_to_clock_t(p->utime),
4214 total = utime + cputime_to_clock_t(p->stime);
4218 * Use CFS's precise accounting:
4220 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4224 do_div(temp, total);
4226 utime = (clock_t)temp;
4228 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4229 return p->prev_utime;
4232 cputime_t task_stime(struct task_struct *p)
4237 * Use CFS's precise accounting. (we subtract utime from
4238 * the total, to make sure the total observed by userspace
4239 * grows monotonically - apps rely on that):
4241 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4242 cputime_to_clock_t(task_utime(p));
4245 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4247 return p->prev_stime;
4251 inline cputime_t task_gtime(struct task_struct *p)
4257 * This function gets called by the timer code, with HZ frequency.
4258 * We call it with interrupts disabled.
4260 * It also gets called by the fork code, when changing the parent's
4263 void scheduler_tick(void)
4265 int cpu = smp_processor_id();
4266 struct rq *rq = cpu_rq(cpu);
4267 struct task_struct *curr = rq->curr;
4271 spin_lock(&rq->lock);
4272 update_rq_clock(rq);
4273 update_cpu_load(rq);
4274 curr->sched_class->task_tick(rq, curr, 0);
4275 spin_unlock(&rq->lock);
4278 rq->idle_at_tick = idle_cpu(cpu);
4279 trigger_load_balance(rq, cpu);
4283 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4284 defined(CONFIG_PREEMPT_TRACER))
4286 static inline unsigned long get_parent_ip(unsigned long addr)
4288 if (in_lock_functions(addr)) {
4289 addr = CALLER_ADDR2;
4290 if (in_lock_functions(addr))
4291 addr = CALLER_ADDR3;
4296 void __kprobes add_preempt_count(int val)
4298 #ifdef CONFIG_DEBUG_PREEMPT
4302 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4305 preempt_count() += val;
4306 #ifdef CONFIG_DEBUG_PREEMPT
4308 * Spinlock count overflowing soon?
4310 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4313 if (preempt_count() == val)
4314 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4316 EXPORT_SYMBOL(add_preempt_count);
4318 void __kprobes sub_preempt_count(int val)
4320 #ifdef CONFIG_DEBUG_PREEMPT
4324 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4327 * Is the spinlock portion underflowing?
4329 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4330 !(preempt_count() & PREEMPT_MASK)))
4334 if (preempt_count() == val)
4335 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4336 preempt_count() -= val;
4338 EXPORT_SYMBOL(sub_preempt_count);
4343 * Print scheduling while atomic bug:
4345 static noinline void __schedule_bug(struct task_struct *prev)
4347 struct pt_regs *regs = get_irq_regs();
4349 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4350 prev->comm, prev->pid, preempt_count());
4352 debug_show_held_locks(prev);
4354 if (irqs_disabled())
4355 print_irqtrace_events(prev);
4364 * Various schedule()-time debugging checks and statistics:
4366 static inline void schedule_debug(struct task_struct *prev)
4369 * Test if we are atomic. Since do_exit() needs to call into
4370 * schedule() atomically, we ignore that path for now.
4371 * Otherwise, whine if we are scheduling when we should not be.
4373 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4374 __schedule_bug(prev);
4376 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4378 schedstat_inc(this_rq(), sched_count);
4379 #ifdef CONFIG_SCHEDSTATS
4380 if (unlikely(prev->lock_depth >= 0)) {
4381 schedstat_inc(this_rq(), bkl_count);
4382 schedstat_inc(prev, sched_info.bkl_count);
4388 * Pick up the highest-prio task:
4390 static inline struct task_struct *
4391 pick_next_task(struct rq *rq, struct task_struct *prev)
4393 const struct sched_class *class;
4394 struct task_struct *p;
4397 * Optimization: we know that if all tasks are in
4398 * the fair class we can call that function directly:
4400 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4401 p = fair_sched_class.pick_next_task(rq);
4406 class = sched_class_highest;
4408 p = class->pick_next_task(rq);
4412 * Will never be NULL as the idle class always
4413 * returns a non-NULL p:
4415 class = class->next;
4420 * schedule() is the main scheduler function.
4422 asmlinkage void __sched schedule(void)
4424 struct task_struct *prev, *next;
4425 unsigned long *switch_count;
4431 cpu = smp_processor_id();
4435 switch_count = &prev->nivcsw;
4437 release_kernel_lock(prev);
4438 need_resched_nonpreemptible:
4440 schedule_debug(prev);
4442 if (sched_feat(HRTICK))
4446 * Do the rq-clock update outside the rq lock:
4448 local_irq_disable();
4449 update_rq_clock(rq);
4450 spin_lock(&rq->lock);
4451 clear_tsk_need_resched(prev);
4453 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4454 if (unlikely(signal_pending_state(prev->state, prev)))
4455 prev->state = TASK_RUNNING;
4457 deactivate_task(rq, prev, 1);
4458 switch_count = &prev->nvcsw;
4462 if (prev->sched_class->pre_schedule)
4463 prev->sched_class->pre_schedule(rq, prev);
4466 if (unlikely(!rq->nr_running))
4467 idle_balance(cpu, rq);
4469 prev->sched_class->put_prev_task(rq, prev);
4470 next = pick_next_task(rq, prev);
4472 if (likely(prev != next)) {
4473 sched_info_switch(prev, next);
4479 context_switch(rq, prev, next); /* unlocks the rq */
4481 * the context switch might have flipped the stack from under
4482 * us, hence refresh the local variables.
4484 cpu = smp_processor_id();
4487 spin_unlock_irq(&rq->lock);
4489 if (unlikely(reacquire_kernel_lock(current) < 0))
4490 goto need_resched_nonpreemptible;
4492 preempt_enable_no_resched();
4493 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4496 EXPORT_SYMBOL(schedule);
4498 #ifdef CONFIG_PREEMPT
4500 * this is the entry point to schedule() from in-kernel preemption
4501 * off of preempt_enable. Kernel preemptions off return from interrupt
4502 * occur there and call schedule directly.
4504 asmlinkage void __sched preempt_schedule(void)
4506 struct thread_info *ti = current_thread_info();
4509 * If there is a non-zero preempt_count or interrupts are disabled,
4510 * we do not want to preempt the current task. Just return..
4512 if (likely(ti->preempt_count || irqs_disabled()))
4516 add_preempt_count(PREEMPT_ACTIVE);
4518 sub_preempt_count(PREEMPT_ACTIVE);
4521 * Check again in case we missed a preemption opportunity
4522 * between schedule and now.
4525 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4527 EXPORT_SYMBOL(preempt_schedule);
4530 * this is the entry point to schedule() from kernel preemption
4531 * off of irq context.
4532 * Note, that this is called and return with irqs disabled. This will
4533 * protect us against recursive calling from irq.
4535 asmlinkage void __sched preempt_schedule_irq(void)
4537 struct thread_info *ti = current_thread_info();
4539 /* Catch callers which need to be fixed */
4540 BUG_ON(ti->preempt_count || !irqs_disabled());
4543 add_preempt_count(PREEMPT_ACTIVE);
4546 local_irq_disable();
4547 sub_preempt_count(PREEMPT_ACTIVE);
4550 * Check again in case we missed a preemption opportunity
4551 * between schedule and now.
4554 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4557 #endif /* CONFIG_PREEMPT */
4559 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4562 return try_to_wake_up(curr->private, mode, sync);
4564 EXPORT_SYMBOL(default_wake_function);
4567 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4568 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4569 * number) then we wake all the non-exclusive tasks and one exclusive task.
4571 * There are circumstances in which we can try to wake a task which has already
4572 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4573 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4575 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4576 int nr_exclusive, int sync, void *key)
4578 wait_queue_t *curr, *next;
4580 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4581 unsigned flags = curr->flags;
4583 if (curr->func(curr, mode, sync, key) &&
4584 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4590 * __wake_up - wake up threads blocked on a waitqueue.
4592 * @mode: which threads
4593 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4594 * @key: is directly passed to the wakeup function
4596 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4597 int nr_exclusive, void *key)
4599 unsigned long flags;
4601 spin_lock_irqsave(&q->lock, flags);
4602 __wake_up_common(q, mode, nr_exclusive, 0, key);
4603 spin_unlock_irqrestore(&q->lock, flags);
4605 EXPORT_SYMBOL(__wake_up);
4608 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4610 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4612 __wake_up_common(q, mode, 1, 0, NULL);
4616 * __wake_up_sync - wake up threads blocked on a waitqueue.
4618 * @mode: which threads
4619 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4621 * The sync wakeup differs that the waker knows that it will schedule
4622 * away soon, so while the target thread will be woken up, it will not
4623 * be migrated to another CPU - ie. the two threads are 'synchronized'
4624 * with each other. This can prevent needless bouncing between CPUs.
4626 * On UP it can prevent extra preemption.
4629 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4631 unsigned long flags;
4637 if (unlikely(!nr_exclusive))
4640 spin_lock_irqsave(&q->lock, flags);
4641 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4642 spin_unlock_irqrestore(&q->lock, flags);
4644 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4647 * complete: - signals a single thread waiting on this completion
4648 * @x: holds the state of this particular completion
4650 * This will wake up a single thread waiting on this completion. Threads will be
4651 * awakened in the same order in which they were queued.
4653 * See also complete_all(), wait_for_completion() and related routines.
4655 void complete(struct completion *x)
4657 unsigned long flags;
4659 spin_lock_irqsave(&x->wait.lock, flags);
4661 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4662 spin_unlock_irqrestore(&x->wait.lock, flags);
4664 EXPORT_SYMBOL(complete);
4667 * complete_all: - signals all threads waiting on this completion
4668 * @x: holds the state of this particular completion
4670 * This will wake up all threads waiting on this particular completion event.
4672 void complete_all(struct completion *x)
4674 unsigned long flags;
4676 spin_lock_irqsave(&x->wait.lock, flags);
4677 x->done += UINT_MAX/2;
4678 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4679 spin_unlock_irqrestore(&x->wait.lock, flags);
4681 EXPORT_SYMBOL(complete_all);
4683 static inline long __sched
4684 do_wait_for_common(struct completion *x, long timeout, int state)
4687 DECLARE_WAITQUEUE(wait, current);
4689 wait.flags |= WQ_FLAG_EXCLUSIVE;
4690 __add_wait_queue_tail(&x->wait, &wait);
4692 if (signal_pending_state(state, current)) {
4693 timeout = -ERESTARTSYS;
4696 __set_current_state(state);
4697 spin_unlock_irq(&x->wait.lock);
4698 timeout = schedule_timeout(timeout);
4699 spin_lock_irq(&x->wait.lock);
4700 } while (!x->done && timeout);
4701 __remove_wait_queue(&x->wait, &wait);
4706 return timeout ?: 1;
4710 wait_for_common(struct completion *x, long timeout, int state)
4714 spin_lock_irq(&x->wait.lock);
4715 timeout = do_wait_for_common(x, timeout, state);
4716 spin_unlock_irq(&x->wait.lock);
4721 * wait_for_completion: - waits for completion of a task
4722 * @x: holds the state of this particular completion
4724 * This waits to be signaled for completion of a specific task. It is NOT
4725 * interruptible and there is no timeout.
4727 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4728 * and interrupt capability. Also see complete().
4730 void __sched wait_for_completion(struct completion *x)
4732 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4734 EXPORT_SYMBOL(wait_for_completion);
4737 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4738 * @x: holds the state of this particular completion
4739 * @timeout: timeout value in jiffies
4741 * This waits for either a completion of a specific task to be signaled or for a
4742 * specified timeout to expire. The timeout is in jiffies. It is not
4745 unsigned long __sched
4746 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4748 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4750 EXPORT_SYMBOL(wait_for_completion_timeout);
4753 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4754 * @x: holds the state of this particular completion
4756 * This waits for completion of a specific task to be signaled. It is
4759 int __sched wait_for_completion_interruptible(struct completion *x)
4761 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4762 if (t == -ERESTARTSYS)
4766 EXPORT_SYMBOL(wait_for_completion_interruptible);
4769 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4770 * @x: holds the state of this particular completion
4771 * @timeout: timeout value in jiffies
4773 * This waits for either a completion of a specific task to be signaled or for a
4774 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4776 unsigned long __sched
4777 wait_for_completion_interruptible_timeout(struct completion *x,
4778 unsigned long timeout)
4780 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4782 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4785 * wait_for_completion_killable: - waits for completion of a task (killable)
4786 * @x: holds the state of this particular completion
4788 * This waits to be signaled for completion of a specific task. It can be
4789 * interrupted by a kill signal.
4791 int __sched wait_for_completion_killable(struct completion *x)
4793 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4794 if (t == -ERESTARTSYS)
4798 EXPORT_SYMBOL(wait_for_completion_killable);
4801 * try_wait_for_completion - try to decrement a completion without blocking
4802 * @x: completion structure
4804 * Returns: 0 if a decrement cannot be done without blocking
4805 * 1 if a decrement succeeded.
4807 * If a completion is being used as a counting completion,
4808 * attempt to decrement the counter without blocking. This
4809 * enables us to avoid waiting if the resource the completion
4810 * is protecting is not available.
4812 bool try_wait_for_completion(struct completion *x)
4816 spin_lock_irq(&x->wait.lock);
4821 spin_unlock_irq(&x->wait.lock);
4824 EXPORT_SYMBOL(try_wait_for_completion);
4827 * completion_done - Test to see if a completion has any waiters
4828 * @x: completion structure
4830 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4831 * 1 if there are no waiters.
4834 bool completion_done(struct completion *x)
4838 spin_lock_irq(&x->wait.lock);
4841 spin_unlock_irq(&x->wait.lock);
4844 EXPORT_SYMBOL(completion_done);
4847 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4849 unsigned long flags;
4852 init_waitqueue_entry(&wait, current);
4854 __set_current_state(state);
4856 spin_lock_irqsave(&q->lock, flags);
4857 __add_wait_queue(q, &wait);
4858 spin_unlock(&q->lock);
4859 timeout = schedule_timeout(timeout);
4860 spin_lock_irq(&q->lock);
4861 __remove_wait_queue(q, &wait);
4862 spin_unlock_irqrestore(&q->lock, flags);
4867 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4869 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4871 EXPORT_SYMBOL(interruptible_sleep_on);
4874 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4876 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4878 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4880 void __sched sleep_on(wait_queue_head_t *q)
4882 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4884 EXPORT_SYMBOL(sleep_on);
4886 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4888 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4890 EXPORT_SYMBOL(sleep_on_timeout);
4892 #ifdef CONFIG_RT_MUTEXES
4895 * rt_mutex_setprio - set the current priority of a task
4897 * @prio: prio value (kernel-internal form)
4899 * This function changes the 'effective' priority of a task. It does
4900 * not touch ->normal_prio like __setscheduler().
4902 * Used by the rt_mutex code to implement priority inheritance logic.
4904 void rt_mutex_setprio(struct task_struct *p, int prio)
4906 unsigned long flags;
4907 int oldprio, on_rq, running;
4909 const struct sched_class *prev_class = p->sched_class;
4911 BUG_ON(prio < 0 || prio > MAX_PRIO);
4913 rq = task_rq_lock(p, &flags);
4914 update_rq_clock(rq);
4917 on_rq = p->se.on_rq;
4918 running = task_current(rq, p);
4920 dequeue_task(rq, p, 0);
4922 p->sched_class->put_prev_task(rq, p);
4925 p->sched_class = &rt_sched_class;
4927 p->sched_class = &fair_sched_class;
4932 p->sched_class->set_curr_task(rq);
4934 enqueue_task(rq, p, 0);
4936 check_class_changed(rq, p, prev_class, oldprio, running);
4938 task_rq_unlock(rq, &flags);
4943 void set_user_nice(struct task_struct *p, long nice)
4945 int old_prio, delta, on_rq;
4946 unsigned long flags;
4949 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4952 * We have to be careful, if called from sys_setpriority(),
4953 * the task might be in the middle of scheduling on another CPU.
4955 rq = task_rq_lock(p, &flags);
4956 update_rq_clock(rq);
4958 * The RT priorities are set via sched_setscheduler(), but we still
4959 * allow the 'normal' nice value to be set - but as expected
4960 * it wont have any effect on scheduling until the task is
4961 * SCHED_FIFO/SCHED_RR:
4963 if (task_has_rt_policy(p)) {
4964 p->static_prio = NICE_TO_PRIO(nice);
4967 on_rq = p->se.on_rq;
4969 dequeue_task(rq, p, 0);
4971 p->static_prio = NICE_TO_PRIO(nice);
4974 p->prio = effective_prio(p);
4975 delta = p->prio - old_prio;
4978 enqueue_task(rq, p, 0);
4980 * If the task increased its priority or is running and
4981 * lowered its priority, then reschedule its CPU:
4983 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4984 resched_task(rq->curr);
4987 task_rq_unlock(rq, &flags);
4989 EXPORT_SYMBOL(set_user_nice);
4992 * can_nice - check if a task can reduce its nice value
4996 int can_nice(const struct task_struct *p, const int nice)
4998 /* convert nice value [19,-20] to rlimit style value [1,40] */
4999 int nice_rlim = 20 - nice;
5001 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5002 capable(CAP_SYS_NICE));
5005 #ifdef __ARCH_WANT_SYS_NICE
5008 * sys_nice - change the priority of the current process.
5009 * @increment: priority increment
5011 * sys_setpriority is a more generic, but much slower function that
5012 * does similar things.
5014 asmlinkage long sys_nice(int increment)
5019 * Setpriority might change our priority at the same moment.
5020 * We don't have to worry. Conceptually one call occurs first
5021 * and we have a single winner.
5023 if (increment < -40)
5028 nice = PRIO_TO_NICE(current->static_prio) + increment;
5034 if (increment < 0 && !can_nice(current, nice))
5037 retval = security_task_setnice(current, nice);
5041 set_user_nice(current, nice);
5048 * task_prio - return the priority value of a given task.
5049 * @p: the task in question.
5051 * This is the priority value as seen by users in /proc.
5052 * RT tasks are offset by -200. Normal tasks are centered
5053 * around 0, value goes from -16 to +15.
5055 int task_prio(const struct task_struct *p)
5057 return p->prio - MAX_RT_PRIO;
5061 * task_nice - return the nice value of a given task.
5062 * @p: the task in question.
5064 int task_nice(const struct task_struct *p)
5066 return TASK_NICE(p);
5068 EXPORT_SYMBOL(task_nice);
5071 * idle_cpu - is a given cpu idle currently?
5072 * @cpu: the processor in question.
5074 int idle_cpu(int cpu)
5076 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5080 * idle_task - return the idle task for a given cpu.
5081 * @cpu: the processor in question.
5083 struct task_struct *idle_task(int cpu)
5085 return cpu_rq(cpu)->idle;
5089 * find_process_by_pid - find a process with a matching PID value.
5090 * @pid: the pid in question.
5092 static struct task_struct *find_process_by_pid(pid_t pid)
5094 return pid ? find_task_by_vpid(pid) : current;
5097 /* Actually do priority change: must hold rq lock. */
5099 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5101 BUG_ON(p->se.on_rq);
5104 switch (p->policy) {
5108 p->sched_class = &fair_sched_class;
5112 p->sched_class = &rt_sched_class;
5116 p->rt_priority = prio;
5117 p->normal_prio = normal_prio(p);
5118 /* we are holding p->pi_lock already */
5119 p->prio = rt_mutex_getprio(p);
5123 static int __sched_setscheduler(struct task_struct *p, int policy,
5124 struct sched_param *param, bool user)
5126 int retval, oldprio, oldpolicy = -1, on_rq, running;
5127 unsigned long flags;
5128 const struct sched_class *prev_class = p->sched_class;
5131 /* may grab non-irq protected spin_locks */
5132 BUG_ON(in_interrupt());
5134 /* double check policy once rq lock held */
5136 policy = oldpolicy = p->policy;
5137 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5138 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5139 policy != SCHED_IDLE)
5142 * Valid priorities for SCHED_FIFO and SCHED_RR are
5143 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5144 * SCHED_BATCH and SCHED_IDLE is 0.
5146 if (param->sched_priority < 0 ||
5147 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5148 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5150 if (rt_policy(policy) != (param->sched_priority != 0))
5154 * Allow unprivileged RT tasks to decrease priority:
5156 if (user && !capable(CAP_SYS_NICE)) {
5157 if (rt_policy(policy)) {
5158 unsigned long rlim_rtprio;
5160 if (!lock_task_sighand(p, &flags))
5162 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5163 unlock_task_sighand(p, &flags);
5165 /* can't set/change the rt policy */
5166 if (policy != p->policy && !rlim_rtprio)
5169 /* can't increase priority */
5170 if (param->sched_priority > p->rt_priority &&
5171 param->sched_priority > rlim_rtprio)
5175 * Like positive nice levels, dont allow tasks to
5176 * move out of SCHED_IDLE either:
5178 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5181 /* can't change other user's priorities */
5182 if ((current->euid != p->euid) &&
5183 (current->euid != p->uid))
5188 #ifdef CONFIG_RT_GROUP_SCHED
5190 * Do not allow realtime tasks into groups that have no runtime
5193 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5194 task_group(p)->rt_bandwidth.rt_runtime == 0)
5198 retval = security_task_setscheduler(p, policy, param);
5204 * make sure no PI-waiters arrive (or leave) while we are
5205 * changing the priority of the task:
5207 spin_lock_irqsave(&p->pi_lock, flags);
5209 * To be able to change p->policy safely, the apropriate
5210 * runqueue lock must be held.
5212 rq = __task_rq_lock(p);
5213 /* recheck policy now with rq lock held */
5214 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5215 policy = oldpolicy = -1;
5216 __task_rq_unlock(rq);
5217 spin_unlock_irqrestore(&p->pi_lock, flags);
5220 update_rq_clock(rq);
5221 on_rq = p->se.on_rq;
5222 running = task_current(rq, p);
5224 deactivate_task(rq, p, 0);
5226 p->sched_class->put_prev_task(rq, p);
5229 __setscheduler(rq, p, policy, param->sched_priority);
5232 p->sched_class->set_curr_task(rq);
5234 activate_task(rq, p, 0);
5236 check_class_changed(rq, p, prev_class, oldprio, running);
5238 __task_rq_unlock(rq);
5239 spin_unlock_irqrestore(&p->pi_lock, flags);
5241 rt_mutex_adjust_pi(p);
5247 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5248 * @p: the task in question.
5249 * @policy: new policy.
5250 * @param: structure containing the new RT priority.
5252 * NOTE that the task may be already dead.
5254 int sched_setscheduler(struct task_struct *p, int policy,
5255 struct sched_param *param)
5257 return __sched_setscheduler(p, policy, param, true);
5259 EXPORT_SYMBOL_GPL(sched_setscheduler);
5262 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5263 * @p: the task in question.
5264 * @policy: new policy.
5265 * @param: structure containing the new RT priority.
5267 * Just like sched_setscheduler, only don't bother checking if the
5268 * current context has permission. For example, this is needed in
5269 * stop_machine(): we create temporary high priority worker threads,
5270 * but our caller might not have that capability.
5272 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5273 struct sched_param *param)
5275 return __sched_setscheduler(p, policy, param, false);
5279 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5281 struct sched_param lparam;
5282 struct task_struct *p;
5285 if (!param || pid < 0)
5287 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5292 p = find_process_by_pid(pid);
5294 retval = sched_setscheduler(p, policy, &lparam);
5301 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5302 * @pid: the pid in question.
5303 * @policy: new policy.
5304 * @param: structure containing the new RT priority.
5307 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5309 /* negative values for policy are not valid */
5313 return do_sched_setscheduler(pid, policy, param);
5317 * sys_sched_setparam - set/change the RT priority of a thread
5318 * @pid: the pid in question.
5319 * @param: structure containing the new RT priority.
5321 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5323 return do_sched_setscheduler(pid, -1, param);
5327 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5328 * @pid: the pid in question.
5330 asmlinkage long sys_sched_getscheduler(pid_t pid)
5332 struct task_struct *p;
5339 read_lock(&tasklist_lock);
5340 p = find_process_by_pid(pid);
5342 retval = security_task_getscheduler(p);
5346 read_unlock(&tasklist_lock);
5351 * sys_sched_getscheduler - get the RT priority of a thread
5352 * @pid: the pid in question.
5353 * @param: structure containing the RT priority.
5355 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5357 struct sched_param lp;
5358 struct task_struct *p;
5361 if (!param || pid < 0)
5364 read_lock(&tasklist_lock);
5365 p = find_process_by_pid(pid);
5370 retval = security_task_getscheduler(p);
5374 lp.sched_priority = p->rt_priority;
5375 read_unlock(&tasklist_lock);
5378 * This one might sleep, we cannot do it with a spinlock held ...
5380 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5385 read_unlock(&tasklist_lock);
5389 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5391 cpumask_t cpus_allowed;
5392 cpumask_t new_mask = *in_mask;
5393 struct task_struct *p;
5397 read_lock(&tasklist_lock);
5399 p = find_process_by_pid(pid);
5401 read_unlock(&tasklist_lock);
5407 * It is not safe to call set_cpus_allowed with the
5408 * tasklist_lock held. We will bump the task_struct's
5409 * usage count and then drop tasklist_lock.
5412 read_unlock(&tasklist_lock);
5415 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5416 !capable(CAP_SYS_NICE))
5419 retval = security_task_setscheduler(p, 0, NULL);
5423 cpuset_cpus_allowed(p, &cpus_allowed);
5424 cpus_and(new_mask, new_mask, cpus_allowed);
5426 retval = set_cpus_allowed_ptr(p, &new_mask);
5429 cpuset_cpus_allowed(p, &cpus_allowed);
5430 if (!cpus_subset(new_mask, cpus_allowed)) {
5432 * We must have raced with a concurrent cpuset
5433 * update. Just reset the cpus_allowed to the
5434 * cpuset's cpus_allowed
5436 new_mask = cpus_allowed;
5446 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5447 cpumask_t *new_mask)
5449 if (len < sizeof(cpumask_t)) {
5450 memset(new_mask, 0, sizeof(cpumask_t));
5451 } else if (len > sizeof(cpumask_t)) {
5452 len = sizeof(cpumask_t);
5454 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5458 * sys_sched_setaffinity - set the cpu affinity of a process
5459 * @pid: pid of the process
5460 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5461 * @user_mask_ptr: user-space pointer to the new cpu mask
5463 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5464 unsigned long __user *user_mask_ptr)
5469 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5473 return sched_setaffinity(pid, &new_mask);
5476 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5478 struct task_struct *p;
5482 read_lock(&tasklist_lock);
5485 p = find_process_by_pid(pid);
5489 retval = security_task_getscheduler(p);
5493 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5496 read_unlock(&tasklist_lock);
5503 * sys_sched_getaffinity - get the cpu affinity of a process
5504 * @pid: pid of the process
5505 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5506 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5508 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5509 unsigned long __user *user_mask_ptr)
5514 if (len < sizeof(cpumask_t))
5517 ret = sched_getaffinity(pid, &mask);
5521 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5524 return sizeof(cpumask_t);
5528 * sys_sched_yield - yield the current processor to other threads.
5530 * This function yields the current CPU to other tasks. If there are no
5531 * other threads running on this CPU then this function will return.
5533 asmlinkage long sys_sched_yield(void)
5535 struct rq *rq = this_rq_lock();
5537 schedstat_inc(rq, yld_count);
5538 current->sched_class->yield_task(rq);
5541 * Since we are going to call schedule() anyway, there's
5542 * no need to preempt or enable interrupts:
5544 __release(rq->lock);
5545 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5546 _raw_spin_unlock(&rq->lock);
5547 preempt_enable_no_resched();
5554 static void __cond_resched(void)
5556 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5557 __might_sleep(__FILE__, __LINE__);
5560 * The BKS might be reacquired before we have dropped
5561 * PREEMPT_ACTIVE, which could trigger a second
5562 * cond_resched() call.
5565 add_preempt_count(PREEMPT_ACTIVE);
5567 sub_preempt_count(PREEMPT_ACTIVE);
5568 } while (need_resched());
5571 int __sched _cond_resched(void)
5573 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5574 system_state == SYSTEM_RUNNING) {
5580 EXPORT_SYMBOL(_cond_resched);
5583 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5584 * call schedule, and on return reacquire the lock.
5586 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5587 * operations here to prevent schedule() from being called twice (once via
5588 * spin_unlock(), once by hand).
5590 int cond_resched_lock(spinlock_t *lock)
5592 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5595 if (spin_needbreak(lock) || resched) {
5597 if (resched && need_resched())
5606 EXPORT_SYMBOL(cond_resched_lock);
5608 int __sched cond_resched_softirq(void)
5610 BUG_ON(!in_softirq());
5612 if (need_resched() && system_state == SYSTEM_RUNNING) {
5620 EXPORT_SYMBOL(cond_resched_softirq);
5623 * yield - yield the current processor to other threads.
5625 * This is a shortcut for kernel-space yielding - it marks the
5626 * thread runnable and calls sys_sched_yield().
5628 void __sched yield(void)
5630 set_current_state(TASK_RUNNING);
5633 EXPORT_SYMBOL(yield);
5636 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5637 * that process accounting knows that this is a task in IO wait state.
5639 * But don't do that if it is a deliberate, throttling IO wait (this task
5640 * has set its backing_dev_info: the queue against which it should throttle)
5642 void __sched io_schedule(void)
5644 struct rq *rq = &__raw_get_cpu_var(runqueues);
5646 delayacct_blkio_start();
5647 atomic_inc(&rq->nr_iowait);
5649 atomic_dec(&rq->nr_iowait);
5650 delayacct_blkio_end();
5652 EXPORT_SYMBOL(io_schedule);
5654 long __sched io_schedule_timeout(long timeout)
5656 struct rq *rq = &__raw_get_cpu_var(runqueues);
5659 delayacct_blkio_start();
5660 atomic_inc(&rq->nr_iowait);
5661 ret = schedule_timeout(timeout);
5662 atomic_dec(&rq->nr_iowait);
5663 delayacct_blkio_end();
5668 * sys_sched_get_priority_max - return maximum RT priority.
5669 * @policy: scheduling class.
5671 * this syscall returns the maximum rt_priority that can be used
5672 * by a given scheduling class.
5674 asmlinkage long sys_sched_get_priority_max(int policy)
5681 ret = MAX_USER_RT_PRIO-1;
5693 * sys_sched_get_priority_min - return minimum RT priority.
5694 * @policy: scheduling class.
5696 * this syscall returns the minimum rt_priority that can be used
5697 * by a given scheduling class.
5699 asmlinkage long sys_sched_get_priority_min(int policy)
5717 * sys_sched_rr_get_interval - return the default timeslice of a process.
5718 * @pid: pid of the process.
5719 * @interval: userspace pointer to the timeslice value.
5721 * this syscall writes the default timeslice value of a given process
5722 * into the user-space timespec buffer. A value of '0' means infinity.
5725 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5727 struct task_struct *p;
5728 unsigned int time_slice;
5736 read_lock(&tasklist_lock);
5737 p = find_process_by_pid(pid);
5741 retval = security_task_getscheduler(p);
5746 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5747 * tasks that are on an otherwise idle runqueue:
5750 if (p->policy == SCHED_RR) {
5751 time_slice = DEF_TIMESLICE;
5752 } else if (p->policy != SCHED_FIFO) {
5753 struct sched_entity *se = &p->se;
5754 unsigned long flags;
5757 rq = task_rq_lock(p, &flags);
5758 if (rq->cfs.load.weight)
5759 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5760 task_rq_unlock(rq, &flags);
5762 read_unlock(&tasklist_lock);
5763 jiffies_to_timespec(time_slice, &t);
5764 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5768 read_unlock(&tasklist_lock);
5772 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5774 void sched_show_task(struct task_struct *p)
5776 unsigned long free = 0;
5779 state = p->state ? __ffs(p->state) + 1 : 0;
5780 printk(KERN_INFO "%-13.13s %c", p->comm,
5781 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5782 #if BITS_PER_LONG == 32
5783 if (state == TASK_RUNNING)
5784 printk(KERN_CONT " running ");
5786 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5788 if (state == TASK_RUNNING)
5789 printk(KERN_CONT " running task ");
5791 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5793 #ifdef CONFIG_DEBUG_STACK_USAGE
5795 unsigned long *n = end_of_stack(p);
5798 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5801 printk(KERN_CONT "%5lu %5d %6d\n", free,
5802 task_pid_nr(p), task_pid_nr(p->real_parent));
5804 show_stack(p, NULL);
5807 void show_state_filter(unsigned long state_filter)
5809 struct task_struct *g, *p;
5811 #if BITS_PER_LONG == 32
5813 " task PC stack pid father\n");
5816 " task PC stack pid father\n");
5818 read_lock(&tasklist_lock);
5819 do_each_thread(g, p) {
5821 * reset the NMI-timeout, listing all files on a slow
5822 * console might take alot of time:
5824 touch_nmi_watchdog();
5825 if (!state_filter || (p->state & state_filter))
5827 } while_each_thread(g, p);
5829 touch_all_softlockup_watchdogs();
5831 #ifdef CONFIG_SCHED_DEBUG
5832 sysrq_sched_debug_show();
5834 read_unlock(&tasklist_lock);
5836 * Only show locks if all tasks are dumped:
5838 if (state_filter == -1)
5839 debug_show_all_locks();
5842 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5844 idle->sched_class = &idle_sched_class;
5848 * init_idle - set up an idle thread for a given CPU
5849 * @idle: task in question
5850 * @cpu: cpu the idle task belongs to
5852 * NOTE: this function does not set the idle thread's NEED_RESCHED
5853 * flag, to make booting more robust.
5855 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5857 struct rq *rq = cpu_rq(cpu);
5858 unsigned long flags;
5861 idle->se.exec_start = sched_clock();
5863 idle->prio = idle->normal_prio = MAX_PRIO;
5864 idle->cpus_allowed = cpumask_of_cpu(cpu);
5865 __set_task_cpu(idle, cpu);
5867 spin_lock_irqsave(&rq->lock, flags);
5868 rq->curr = rq->idle = idle;
5869 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5872 spin_unlock_irqrestore(&rq->lock, flags);
5874 /* Set the preempt count _outside_ the spinlocks! */
5875 #if defined(CONFIG_PREEMPT)
5876 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5878 task_thread_info(idle)->preempt_count = 0;
5881 * The idle tasks have their own, simple scheduling class:
5883 idle->sched_class = &idle_sched_class;
5887 * In a system that switches off the HZ timer nohz_cpu_mask
5888 * indicates which cpus entered this state. This is used
5889 * in the rcu update to wait only for active cpus. For system
5890 * which do not switch off the HZ timer nohz_cpu_mask should
5891 * always be CPU_MASK_NONE.
5893 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5896 * Increase the granularity value when there are more CPUs,
5897 * because with more CPUs the 'effective latency' as visible
5898 * to users decreases. But the relationship is not linear,
5899 * so pick a second-best guess by going with the log2 of the
5902 * This idea comes from the SD scheduler of Con Kolivas:
5904 static inline void sched_init_granularity(void)
5906 unsigned int factor = 1 + ilog2(num_online_cpus());
5907 const unsigned long limit = 200000000;
5909 sysctl_sched_min_granularity *= factor;
5910 if (sysctl_sched_min_granularity > limit)
5911 sysctl_sched_min_granularity = limit;
5913 sysctl_sched_latency *= factor;
5914 if (sysctl_sched_latency > limit)
5915 sysctl_sched_latency = limit;
5917 sysctl_sched_wakeup_granularity *= factor;
5919 sysctl_sched_shares_ratelimit *= factor;
5924 * This is how migration works:
5926 * 1) we queue a struct migration_req structure in the source CPU's
5927 * runqueue and wake up that CPU's migration thread.
5928 * 2) we down() the locked semaphore => thread blocks.
5929 * 3) migration thread wakes up (implicitly it forces the migrated
5930 * thread off the CPU)
5931 * 4) it gets the migration request and checks whether the migrated
5932 * task is still in the wrong runqueue.
5933 * 5) if it's in the wrong runqueue then the migration thread removes
5934 * it and puts it into the right queue.
5935 * 6) migration thread up()s the semaphore.
5936 * 7) we wake up and the migration is done.
5940 * Change a given task's CPU affinity. Migrate the thread to a
5941 * proper CPU and schedule it away if the CPU it's executing on
5942 * is removed from the allowed bitmask.
5944 * NOTE: the caller must have a valid reference to the task, the
5945 * task must not exit() & deallocate itself prematurely. The
5946 * call is not atomic; no spinlocks may be held.
5948 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5950 struct migration_req req;
5951 unsigned long flags;
5955 rq = task_rq_lock(p, &flags);
5956 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5961 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5962 !cpus_equal(p->cpus_allowed, *new_mask))) {
5967 if (p->sched_class->set_cpus_allowed)
5968 p->sched_class->set_cpus_allowed(p, new_mask);
5970 p->cpus_allowed = *new_mask;
5971 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5974 /* Can the task run on the task's current CPU? If so, we're done */
5975 if (cpu_isset(task_cpu(p), *new_mask))
5978 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5979 /* Need help from migration thread: drop lock and wait. */
5980 task_rq_unlock(rq, &flags);
5981 wake_up_process(rq->migration_thread);
5982 wait_for_completion(&req.done);
5983 tlb_migrate_finish(p->mm);
5987 task_rq_unlock(rq, &flags);
5991 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5994 * Move (not current) task off this cpu, onto dest cpu. We're doing
5995 * this because either it can't run here any more (set_cpus_allowed()
5996 * away from this CPU, or CPU going down), or because we're
5997 * attempting to rebalance this task on exec (sched_exec).
5999 * So we race with normal scheduler movements, but that's OK, as long
6000 * as the task is no longer on this CPU.
6002 * Returns non-zero if task was successfully migrated.
6004 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6006 struct rq *rq_dest, *rq_src;
6009 if (unlikely(!cpu_active(dest_cpu)))
6012 rq_src = cpu_rq(src_cpu);
6013 rq_dest = cpu_rq(dest_cpu);
6015 double_rq_lock(rq_src, rq_dest);
6016 /* Already moved. */
6017 if (task_cpu(p) != src_cpu)
6019 /* Affinity changed (again). */
6020 if (!cpu_isset(dest_cpu, p->cpus_allowed))
6023 on_rq = p->se.on_rq;
6025 deactivate_task(rq_src, p, 0);
6027 set_task_cpu(p, dest_cpu);
6029 activate_task(rq_dest, p, 0);
6030 check_preempt_curr(rq_dest, p, 0);
6035 double_rq_unlock(rq_src, rq_dest);
6040 * migration_thread - this is a highprio system thread that performs
6041 * thread migration by bumping thread off CPU then 'pushing' onto
6044 static int migration_thread(void *data)
6046 int cpu = (long)data;
6050 BUG_ON(rq->migration_thread != current);
6052 set_current_state(TASK_INTERRUPTIBLE);
6053 while (!kthread_should_stop()) {
6054 struct migration_req *req;
6055 struct list_head *head;
6057 spin_lock_irq(&rq->lock);
6059 if (cpu_is_offline(cpu)) {
6060 spin_unlock_irq(&rq->lock);
6064 if (rq->active_balance) {
6065 active_load_balance(rq, cpu);
6066 rq->active_balance = 0;
6069 head = &rq->migration_queue;
6071 if (list_empty(head)) {
6072 spin_unlock_irq(&rq->lock);
6074 set_current_state(TASK_INTERRUPTIBLE);
6077 req = list_entry(head->next, struct migration_req, list);
6078 list_del_init(head->next);
6080 spin_unlock(&rq->lock);
6081 __migrate_task(req->task, cpu, req->dest_cpu);
6084 complete(&req->done);
6086 __set_current_state(TASK_RUNNING);
6090 /* Wait for kthread_stop */
6091 set_current_state(TASK_INTERRUPTIBLE);
6092 while (!kthread_should_stop()) {
6094 set_current_state(TASK_INTERRUPTIBLE);
6096 __set_current_state(TASK_RUNNING);
6100 #ifdef CONFIG_HOTPLUG_CPU
6102 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6106 local_irq_disable();
6107 ret = __migrate_task(p, src_cpu, dest_cpu);
6113 * Figure out where task on dead CPU should go, use force if necessary.
6114 * NOTE: interrupts should be disabled by the caller
6116 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6118 unsigned long flags;
6125 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6126 cpus_and(mask, mask, p->cpus_allowed);
6127 dest_cpu = any_online_cpu(mask);
6129 /* On any allowed CPU? */
6130 if (dest_cpu >= nr_cpu_ids)
6131 dest_cpu = any_online_cpu(p->cpus_allowed);
6133 /* No more Mr. Nice Guy. */
6134 if (dest_cpu >= nr_cpu_ids) {
6135 cpumask_t cpus_allowed;
6137 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6139 * Try to stay on the same cpuset, where the
6140 * current cpuset may be a subset of all cpus.
6141 * The cpuset_cpus_allowed_locked() variant of
6142 * cpuset_cpus_allowed() will not block. It must be
6143 * called within calls to cpuset_lock/cpuset_unlock.
6145 rq = task_rq_lock(p, &flags);
6146 p->cpus_allowed = cpus_allowed;
6147 dest_cpu = any_online_cpu(p->cpus_allowed);
6148 task_rq_unlock(rq, &flags);
6151 * Don't tell them about moving exiting tasks or
6152 * kernel threads (both mm NULL), since they never
6155 if (p->mm && printk_ratelimit()) {
6156 printk(KERN_INFO "process %d (%s) no "
6157 "longer affine to cpu%d\n",
6158 task_pid_nr(p), p->comm, dead_cpu);
6161 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6165 * While a dead CPU has no uninterruptible tasks queued at this point,
6166 * it might still have a nonzero ->nr_uninterruptible counter, because
6167 * for performance reasons the counter is not stricly tracking tasks to
6168 * their home CPUs. So we just add the counter to another CPU's counter,
6169 * to keep the global sum constant after CPU-down:
6171 static void migrate_nr_uninterruptible(struct rq *rq_src)
6173 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6174 unsigned long flags;
6176 local_irq_save(flags);
6177 double_rq_lock(rq_src, rq_dest);
6178 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6179 rq_src->nr_uninterruptible = 0;
6180 double_rq_unlock(rq_src, rq_dest);
6181 local_irq_restore(flags);
6184 /* Run through task list and migrate tasks from the dead cpu. */
6185 static void migrate_live_tasks(int src_cpu)
6187 struct task_struct *p, *t;
6189 read_lock(&tasklist_lock);
6191 do_each_thread(t, p) {
6195 if (task_cpu(p) == src_cpu)
6196 move_task_off_dead_cpu(src_cpu, p);
6197 } while_each_thread(t, p);
6199 read_unlock(&tasklist_lock);
6203 * Schedules idle task to be the next runnable task on current CPU.
6204 * It does so by boosting its priority to highest possible.
6205 * Used by CPU offline code.
6207 void sched_idle_next(void)
6209 int this_cpu = smp_processor_id();
6210 struct rq *rq = cpu_rq(this_cpu);
6211 struct task_struct *p = rq->idle;
6212 unsigned long flags;
6214 /* cpu has to be offline */
6215 BUG_ON(cpu_online(this_cpu));
6218 * Strictly not necessary since rest of the CPUs are stopped by now
6219 * and interrupts disabled on the current cpu.
6221 spin_lock_irqsave(&rq->lock, flags);
6223 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6225 update_rq_clock(rq);
6226 activate_task(rq, p, 0);
6228 spin_unlock_irqrestore(&rq->lock, flags);
6232 * Ensures that the idle task is using init_mm right before its cpu goes
6235 void idle_task_exit(void)
6237 struct mm_struct *mm = current->active_mm;
6239 BUG_ON(cpu_online(smp_processor_id()));
6242 switch_mm(mm, &init_mm, current);
6246 /* called under rq->lock with disabled interrupts */
6247 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6249 struct rq *rq = cpu_rq(dead_cpu);
6251 /* Must be exiting, otherwise would be on tasklist. */
6252 BUG_ON(!p->exit_state);
6254 /* Cannot have done final schedule yet: would have vanished. */
6255 BUG_ON(p->state == TASK_DEAD);
6260 * Drop lock around migration; if someone else moves it,
6261 * that's OK. No task can be added to this CPU, so iteration is
6264 spin_unlock_irq(&rq->lock);
6265 move_task_off_dead_cpu(dead_cpu, p);
6266 spin_lock_irq(&rq->lock);
6271 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6272 static void migrate_dead_tasks(unsigned int dead_cpu)
6274 struct rq *rq = cpu_rq(dead_cpu);
6275 struct task_struct *next;
6278 if (!rq->nr_running)
6280 update_rq_clock(rq);
6281 next = pick_next_task(rq, rq->curr);
6284 next->sched_class->put_prev_task(rq, next);
6285 migrate_dead(dead_cpu, next);
6289 #endif /* CONFIG_HOTPLUG_CPU */
6291 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6293 static struct ctl_table sd_ctl_dir[] = {
6295 .procname = "sched_domain",
6301 static struct ctl_table sd_ctl_root[] = {
6303 .ctl_name = CTL_KERN,
6304 .procname = "kernel",
6306 .child = sd_ctl_dir,
6311 static struct ctl_table *sd_alloc_ctl_entry(int n)
6313 struct ctl_table *entry =
6314 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6319 static void sd_free_ctl_entry(struct ctl_table **tablep)
6321 struct ctl_table *entry;
6324 * In the intermediate directories, both the child directory and
6325 * procname are dynamically allocated and could fail but the mode
6326 * will always be set. In the lowest directory the names are
6327 * static strings and all have proc handlers.
6329 for (entry = *tablep; entry->mode; entry++) {
6331 sd_free_ctl_entry(&entry->child);
6332 if (entry->proc_handler == NULL)
6333 kfree(entry->procname);
6341 set_table_entry(struct ctl_table *entry,
6342 const char *procname, void *data, int maxlen,
6343 mode_t mode, proc_handler *proc_handler)
6345 entry->procname = procname;
6347 entry->maxlen = maxlen;
6349 entry->proc_handler = proc_handler;
6352 static struct ctl_table *
6353 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6355 struct ctl_table *table = sd_alloc_ctl_entry(13);
6360 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6361 sizeof(long), 0644, proc_doulongvec_minmax);
6362 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6363 sizeof(long), 0644, proc_doulongvec_minmax);
6364 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6365 sizeof(int), 0644, proc_dointvec_minmax);
6366 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6367 sizeof(int), 0644, proc_dointvec_minmax);
6368 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6369 sizeof(int), 0644, proc_dointvec_minmax);
6370 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6371 sizeof(int), 0644, proc_dointvec_minmax);
6372 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6373 sizeof(int), 0644, proc_dointvec_minmax);
6374 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6375 sizeof(int), 0644, proc_dointvec_minmax);
6376 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6377 sizeof(int), 0644, proc_dointvec_minmax);
6378 set_table_entry(&table[9], "cache_nice_tries",
6379 &sd->cache_nice_tries,
6380 sizeof(int), 0644, proc_dointvec_minmax);
6381 set_table_entry(&table[10], "flags", &sd->flags,
6382 sizeof(int), 0644, proc_dointvec_minmax);
6383 set_table_entry(&table[11], "name", sd->name,
6384 CORENAME_MAX_SIZE, 0444, proc_dostring);
6385 /* &table[12] is terminator */
6390 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6392 struct ctl_table *entry, *table;
6393 struct sched_domain *sd;
6394 int domain_num = 0, i;
6397 for_each_domain(cpu, sd)
6399 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6404 for_each_domain(cpu, sd) {
6405 snprintf(buf, 32, "domain%d", i);
6406 entry->procname = kstrdup(buf, GFP_KERNEL);
6408 entry->child = sd_alloc_ctl_domain_table(sd);
6415 static struct ctl_table_header *sd_sysctl_header;
6416 static void register_sched_domain_sysctl(void)
6418 int i, cpu_num = num_online_cpus();
6419 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6422 WARN_ON(sd_ctl_dir[0].child);
6423 sd_ctl_dir[0].child = entry;
6428 for_each_online_cpu(i) {
6429 snprintf(buf, 32, "cpu%d", i);
6430 entry->procname = kstrdup(buf, GFP_KERNEL);
6432 entry->child = sd_alloc_ctl_cpu_table(i);
6436 WARN_ON(sd_sysctl_header);
6437 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6440 /* may be called multiple times per register */
6441 static void unregister_sched_domain_sysctl(void)
6443 if (sd_sysctl_header)
6444 unregister_sysctl_table(sd_sysctl_header);
6445 sd_sysctl_header = NULL;
6446 if (sd_ctl_dir[0].child)
6447 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6450 static void register_sched_domain_sysctl(void)
6453 static void unregister_sched_domain_sysctl(void)
6458 static void set_rq_online(struct rq *rq)
6461 const struct sched_class *class;
6463 cpu_set(rq->cpu, rq->rd->online);
6466 for_each_class(class) {
6467 if (class->rq_online)
6468 class->rq_online(rq);
6473 static void set_rq_offline(struct rq *rq)
6476 const struct sched_class *class;
6478 for_each_class(class) {
6479 if (class->rq_offline)
6480 class->rq_offline(rq);
6483 cpu_clear(rq->cpu, rq->rd->online);
6489 * migration_call - callback that gets triggered when a CPU is added.
6490 * Here we can start up the necessary migration thread for the new CPU.
6492 static int __cpuinit
6493 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6495 struct task_struct *p;
6496 int cpu = (long)hcpu;
6497 unsigned long flags;
6502 case CPU_UP_PREPARE:
6503 case CPU_UP_PREPARE_FROZEN:
6504 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6507 kthread_bind(p, cpu);
6508 /* Must be high prio: stop_machine expects to yield to it. */
6509 rq = task_rq_lock(p, &flags);
6510 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6511 task_rq_unlock(rq, &flags);
6512 cpu_rq(cpu)->migration_thread = p;
6516 case CPU_ONLINE_FROZEN:
6517 /* Strictly unnecessary, as first user will wake it. */
6518 wake_up_process(cpu_rq(cpu)->migration_thread);
6520 /* Update our root-domain */
6522 spin_lock_irqsave(&rq->lock, flags);
6524 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6528 spin_unlock_irqrestore(&rq->lock, flags);
6531 #ifdef CONFIG_HOTPLUG_CPU
6532 case CPU_UP_CANCELED:
6533 case CPU_UP_CANCELED_FROZEN:
6534 if (!cpu_rq(cpu)->migration_thread)
6536 /* Unbind it from offline cpu so it can run. Fall thru. */
6537 kthread_bind(cpu_rq(cpu)->migration_thread,
6538 any_online_cpu(cpu_online_map));
6539 kthread_stop(cpu_rq(cpu)->migration_thread);
6540 cpu_rq(cpu)->migration_thread = NULL;
6544 case CPU_DEAD_FROZEN:
6545 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6546 migrate_live_tasks(cpu);
6548 kthread_stop(rq->migration_thread);
6549 rq->migration_thread = NULL;
6550 /* Idle task back to normal (off runqueue, low prio) */
6551 spin_lock_irq(&rq->lock);
6552 update_rq_clock(rq);
6553 deactivate_task(rq, rq->idle, 0);
6554 rq->idle->static_prio = MAX_PRIO;
6555 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6556 rq->idle->sched_class = &idle_sched_class;
6557 migrate_dead_tasks(cpu);
6558 spin_unlock_irq(&rq->lock);
6560 migrate_nr_uninterruptible(rq);
6561 BUG_ON(rq->nr_running != 0);
6564 * No need to migrate the tasks: it was best-effort if
6565 * they didn't take sched_hotcpu_mutex. Just wake up
6568 spin_lock_irq(&rq->lock);
6569 while (!list_empty(&rq->migration_queue)) {
6570 struct migration_req *req;
6572 req = list_entry(rq->migration_queue.next,
6573 struct migration_req, list);
6574 list_del_init(&req->list);
6575 complete(&req->done);
6577 spin_unlock_irq(&rq->lock);
6581 case CPU_DYING_FROZEN:
6582 /* Update our root-domain */
6584 spin_lock_irqsave(&rq->lock, flags);
6586 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6589 spin_unlock_irqrestore(&rq->lock, flags);
6596 /* Register at highest priority so that task migration (migrate_all_tasks)
6597 * happens before everything else.
6599 static struct notifier_block __cpuinitdata migration_notifier = {
6600 .notifier_call = migration_call,
6604 static int __init migration_init(void)
6606 void *cpu = (void *)(long)smp_processor_id();
6609 /* Start one for the boot CPU: */
6610 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6611 BUG_ON(err == NOTIFY_BAD);
6612 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6613 register_cpu_notifier(&migration_notifier);
6617 early_initcall(migration_init);
6622 #ifdef CONFIG_SCHED_DEBUG
6624 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6637 case SD_LV_ALLNODES:
6646 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6647 cpumask_t *groupmask)
6649 struct sched_group *group = sd->groups;
6652 cpulist_scnprintf(str, sizeof(str), sd->span);
6653 cpus_clear(*groupmask);
6655 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6657 if (!(sd->flags & SD_LOAD_BALANCE)) {
6658 printk("does not load-balance\n");
6660 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6665 printk(KERN_CONT "span %s level %s\n",
6666 str, sd_level_to_string(sd->level));
6668 if (!cpu_isset(cpu, sd->span)) {
6669 printk(KERN_ERR "ERROR: domain->span does not contain "
6672 if (!cpu_isset(cpu, group->cpumask)) {
6673 printk(KERN_ERR "ERROR: domain->groups does not contain"
6677 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6681 printk(KERN_ERR "ERROR: group is NULL\n");
6685 if (!group->__cpu_power) {
6686 printk(KERN_CONT "\n");
6687 printk(KERN_ERR "ERROR: domain->cpu_power not "
6692 if (!cpus_weight(group->cpumask)) {
6693 printk(KERN_CONT "\n");
6694 printk(KERN_ERR "ERROR: empty group\n");
6698 if (cpus_intersects(*groupmask, group->cpumask)) {
6699 printk(KERN_CONT "\n");
6700 printk(KERN_ERR "ERROR: repeated CPUs\n");
6704 cpus_or(*groupmask, *groupmask, group->cpumask);
6706 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6707 printk(KERN_CONT " %s", str);
6709 group = group->next;
6710 } while (group != sd->groups);
6711 printk(KERN_CONT "\n");
6713 if (!cpus_equal(sd->span, *groupmask))
6714 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6716 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6717 printk(KERN_ERR "ERROR: parent span is not a superset "
6718 "of domain->span\n");
6722 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6724 cpumask_t *groupmask;
6728 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6732 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6734 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6736 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6741 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6750 #else /* !CONFIG_SCHED_DEBUG */
6751 # define sched_domain_debug(sd, cpu) do { } while (0)
6752 #endif /* CONFIG_SCHED_DEBUG */
6754 static int sd_degenerate(struct sched_domain *sd)
6756 if (cpus_weight(sd->span) == 1)
6759 /* Following flags need at least 2 groups */
6760 if (sd->flags & (SD_LOAD_BALANCE |
6761 SD_BALANCE_NEWIDLE |
6765 SD_SHARE_PKG_RESOURCES)) {
6766 if (sd->groups != sd->groups->next)
6770 /* Following flags don't use groups */
6771 if (sd->flags & (SD_WAKE_IDLE |
6780 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6782 unsigned long cflags = sd->flags, pflags = parent->flags;
6784 if (sd_degenerate(parent))
6787 if (!cpus_equal(sd->span, parent->span))
6790 /* Does parent contain flags not in child? */
6791 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6792 if (cflags & SD_WAKE_AFFINE)
6793 pflags &= ~SD_WAKE_BALANCE;
6794 /* Flags needing groups don't count if only 1 group in parent */
6795 if (parent->groups == parent->groups->next) {
6796 pflags &= ~(SD_LOAD_BALANCE |
6797 SD_BALANCE_NEWIDLE |
6801 SD_SHARE_PKG_RESOURCES);
6803 if (~cflags & pflags)
6809 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6811 unsigned long flags;
6813 spin_lock_irqsave(&rq->lock, flags);
6816 struct root_domain *old_rd = rq->rd;
6818 if (cpu_isset(rq->cpu, old_rd->online))
6821 cpu_clear(rq->cpu, old_rd->span);
6823 if (atomic_dec_and_test(&old_rd->refcount))
6827 atomic_inc(&rd->refcount);
6830 cpu_set(rq->cpu, rd->span);
6831 if (cpu_isset(rq->cpu, cpu_online_map))
6834 spin_unlock_irqrestore(&rq->lock, flags);
6837 static void init_rootdomain(struct root_domain *rd)
6839 memset(rd, 0, sizeof(*rd));
6841 cpus_clear(rd->span);
6842 cpus_clear(rd->online);
6844 cpupri_init(&rd->cpupri);
6847 static void init_defrootdomain(void)
6849 init_rootdomain(&def_root_domain);
6850 atomic_set(&def_root_domain.refcount, 1);
6853 static struct root_domain *alloc_rootdomain(void)
6855 struct root_domain *rd;
6857 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6861 init_rootdomain(rd);
6867 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6868 * hold the hotplug lock.
6871 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6873 struct rq *rq = cpu_rq(cpu);
6874 struct sched_domain *tmp;
6876 /* Remove the sched domains which do not contribute to scheduling. */
6877 for (tmp = sd; tmp; tmp = tmp->parent) {
6878 struct sched_domain *parent = tmp->parent;
6881 if (sd_parent_degenerate(tmp, parent)) {
6882 tmp->parent = parent->parent;
6884 parent->parent->child = tmp;
6888 if (sd && sd_degenerate(sd)) {
6894 sched_domain_debug(sd, cpu);
6896 rq_attach_root(rq, rd);
6897 rcu_assign_pointer(rq->sd, sd);
6900 /* cpus with isolated domains */
6901 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6903 /* Setup the mask of cpus configured for isolated domains */
6904 static int __init isolated_cpu_setup(char *str)
6906 static int __initdata ints[NR_CPUS];
6909 str = get_options(str, ARRAY_SIZE(ints), ints);
6910 cpus_clear(cpu_isolated_map);
6911 for (i = 1; i <= ints[0]; i++)
6912 if (ints[i] < NR_CPUS)
6913 cpu_set(ints[i], cpu_isolated_map);
6917 __setup("isolcpus=", isolated_cpu_setup);
6920 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6921 * to a function which identifies what group(along with sched group) a CPU
6922 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6923 * (due to the fact that we keep track of groups covered with a cpumask_t).
6925 * init_sched_build_groups will build a circular linked list of the groups
6926 * covered by the given span, and will set each group's ->cpumask correctly,
6927 * and ->cpu_power to 0.
6930 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6931 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6932 struct sched_group **sg,
6933 cpumask_t *tmpmask),
6934 cpumask_t *covered, cpumask_t *tmpmask)
6936 struct sched_group *first = NULL, *last = NULL;
6939 cpus_clear(*covered);
6941 for_each_cpu_mask_nr(i, *span) {
6942 struct sched_group *sg;
6943 int group = group_fn(i, cpu_map, &sg, tmpmask);
6946 if (cpu_isset(i, *covered))
6949 cpus_clear(sg->cpumask);
6950 sg->__cpu_power = 0;
6952 for_each_cpu_mask_nr(j, *span) {
6953 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6956 cpu_set(j, *covered);
6957 cpu_set(j, sg->cpumask);
6968 #define SD_NODES_PER_DOMAIN 16
6973 * find_next_best_node - find the next node to include in a sched_domain
6974 * @node: node whose sched_domain we're building
6975 * @used_nodes: nodes already in the sched_domain
6977 * Find the next node to include in a given scheduling domain. Simply
6978 * finds the closest node not already in the @used_nodes map.
6980 * Should use nodemask_t.
6982 static int find_next_best_node(int node, nodemask_t *used_nodes)
6984 int i, n, val, min_val, best_node = 0;
6988 for (i = 0; i < nr_node_ids; i++) {
6989 /* Start at @node */
6990 n = (node + i) % nr_node_ids;
6992 if (!nr_cpus_node(n))
6995 /* Skip already used nodes */
6996 if (node_isset(n, *used_nodes))
6999 /* Simple min distance search */
7000 val = node_distance(node, n);
7002 if (val < min_val) {
7008 node_set(best_node, *used_nodes);
7013 * sched_domain_node_span - get a cpumask for a node's sched_domain
7014 * @node: node whose cpumask we're constructing
7015 * @span: resulting cpumask
7017 * Given a node, construct a good cpumask for its sched_domain to span. It
7018 * should be one that prevents unnecessary balancing, but also spreads tasks
7021 static void sched_domain_node_span(int node, cpumask_t *span)
7023 nodemask_t used_nodes;
7024 node_to_cpumask_ptr(nodemask, node);
7028 nodes_clear(used_nodes);
7030 cpus_or(*span, *span, *nodemask);
7031 node_set(node, used_nodes);
7033 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7034 int next_node = find_next_best_node(node, &used_nodes);
7036 node_to_cpumask_ptr_next(nodemask, next_node);
7037 cpus_or(*span, *span, *nodemask);
7040 #endif /* CONFIG_NUMA */
7042 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7045 * SMT sched-domains:
7047 #ifdef CONFIG_SCHED_SMT
7048 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
7049 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
7052 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7056 *sg = &per_cpu(sched_group_cpus, cpu);
7059 #endif /* CONFIG_SCHED_SMT */
7062 * multi-core sched-domains:
7064 #ifdef CONFIG_SCHED_MC
7065 static DEFINE_PER_CPU(struct sched_domain, core_domains);
7066 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
7067 #endif /* CONFIG_SCHED_MC */
7069 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7071 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7076 *mask = per_cpu(cpu_sibling_map, cpu);
7077 cpus_and(*mask, *mask, *cpu_map);
7078 group = first_cpu(*mask);
7080 *sg = &per_cpu(sched_group_core, group);
7083 #elif defined(CONFIG_SCHED_MC)
7085 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7089 *sg = &per_cpu(sched_group_core, cpu);
7094 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
7095 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
7098 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7102 #ifdef CONFIG_SCHED_MC
7103 *mask = cpu_coregroup_map(cpu);
7104 cpus_and(*mask, *mask, *cpu_map);
7105 group = first_cpu(*mask);
7106 #elif defined(CONFIG_SCHED_SMT)
7107 *mask = per_cpu(cpu_sibling_map, cpu);
7108 cpus_and(*mask, *mask, *cpu_map);
7109 group = first_cpu(*mask);
7114 *sg = &per_cpu(sched_group_phys, group);
7120 * The init_sched_build_groups can't handle what we want to do with node
7121 * groups, so roll our own. Now each node has its own list of groups which
7122 * gets dynamically allocated.
7124 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7125 static struct sched_group ***sched_group_nodes_bycpu;
7127 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7128 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7130 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7131 struct sched_group **sg, cpumask_t *nodemask)
7135 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7136 cpus_and(*nodemask, *nodemask, *cpu_map);
7137 group = first_cpu(*nodemask);
7140 *sg = &per_cpu(sched_group_allnodes, group);
7144 static void init_numa_sched_groups_power(struct sched_group *group_head)
7146 struct sched_group *sg = group_head;
7152 for_each_cpu_mask_nr(j, sg->cpumask) {
7153 struct sched_domain *sd;
7155 sd = &per_cpu(phys_domains, j);
7156 if (j != first_cpu(sd->groups->cpumask)) {
7158 * Only add "power" once for each
7164 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7167 } while (sg != group_head);
7169 #endif /* CONFIG_NUMA */
7172 /* Free memory allocated for various sched_group structures */
7173 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7177 for_each_cpu_mask_nr(cpu, *cpu_map) {
7178 struct sched_group **sched_group_nodes
7179 = sched_group_nodes_bycpu[cpu];
7181 if (!sched_group_nodes)
7184 for (i = 0; i < nr_node_ids; i++) {
7185 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7187 *nodemask = node_to_cpumask(i);
7188 cpus_and(*nodemask, *nodemask, *cpu_map);
7189 if (cpus_empty(*nodemask))
7199 if (oldsg != sched_group_nodes[i])
7202 kfree(sched_group_nodes);
7203 sched_group_nodes_bycpu[cpu] = NULL;
7206 #else /* !CONFIG_NUMA */
7207 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7210 #endif /* CONFIG_NUMA */
7213 * Initialize sched groups cpu_power.
7215 * cpu_power indicates the capacity of sched group, which is used while
7216 * distributing the load between different sched groups in a sched domain.
7217 * Typically cpu_power for all the groups in a sched domain will be same unless
7218 * there are asymmetries in the topology. If there are asymmetries, group
7219 * having more cpu_power will pickup more load compared to the group having
7222 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7223 * the maximum number of tasks a group can handle in the presence of other idle
7224 * or lightly loaded groups in the same sched domain.
7226 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7228 struct sched_domain *child;
7229 struct sched_group *group;
7231 WARN_ON(!sd || !sd->groups);
7233 if (cpu != first_cpu(sd->groups->cpumask))
7238 sd->groups->__cpu_power = 0;
7241 * For perf policy, if the groups in child domain share resources
7242 * (for example cores sharing some portions of the cache hierarchy
7243 * or SMT), then set this domain groups cpu_power such that each group
7244 * can handle only one task, when there are other idle groups in the
7245 * same sched domain.
7247 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7249 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7250 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7255 * add cpu_power of each child group to this groups cpu_power
7257 group = child->groups;
7259 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7260 group = group->next;
7261 } while (group != child->groups);
7265 * Initializers for schedule domains
7266 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7269 #ifdef CONFIG_SCHED_DEBUG
7270 # define SD_INIT_NAME(sd, type) sd->name = #type
7272 # define SD_INIT_NAME(sd, type) do { } while (0)
7275 #define SD_INIT(sd, type) sd_init_##type(sd)
7277 #define SD_INIT_FUNC(type) \
7278 static noinline void sd_init_##type(struct sched_domain *sd) \
7280 memset(sd, 0, sizeof(*sd)); \
7281 *sd = SD_##type##_INIT; \
7282 sd->level = SD_LV_##type; \
7283 SD_INIT_NAME(sd, type); \
7288 SD_INIT_FUNC(ALLNODES)
7291 #ifdef CONFIG_SCHED_SMT
7292 SD_INIT_FUNC(SIBLING)
7294 #ifdef CONFIG_SCHED_MC
7299 * To minimize stack usage kmalloc room for cpumasks and share the
7300 * space as the usage in build_sched_domains() dictates. Used only
7301 * if the amount of space is significant.
7304 cpumask_t tmpmask; /* make this one first */
7307 cpumask_t this_sibling_map;
7308 cpumask_t this_core_map;
7310 cpumask_t send_covered;
7313 cpumask_t domainspan;
7315 cpumask_t notcovered;
7320 #define SCHED_CPUMASK_ALLOC 1
7321 #define SCHED_CPUMASK_FREE(v) kfree(v)
7322 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7324 #define SCHED_CPUMASK_ALLOC 0
7325 #define SCHED_CPUMASK_FREE(v)
7326 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7329 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7330 ((unsigned long)(a) + offsetof(struct allmasks, v))
7332 static int default_relax_domain_level = -1;
7334 static int __init setup_relax_domain_level(char *str)
7338 val = simple_strtoul(str, NULL, 0);
7339 if (val < SD_LV_MAX)
7340 default_relax_domain_level = val;
7344 __setup("relax_domain_level=", setup_relax_domain_level);
7346 static void set_domain_attribute(struct sched_domain *sd,
7347 struct sched_domain_attr *attr)
7351 if (!attr || attr->relax_domain_level < 0) {
7352 if (default_relax_domain_level < 0)
7355 request = default_relax_domain_level;
7357 request = attr->relax_domain_level;
7358 if (request < sd->level) {
7359 /* turn off idle balance on this domain */
7360 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7362 /* turn on idle balance on this domain */
7363 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7368 * Build sched domains for a given set of cpus and attach the sched domains
7369 * to the individual cpus
7371 static int __build_sched_domains(const cpumask_t *cpu_map,
7372 struct sched_domain_attr *attr)
7375 struct root_domain *rd;
7376 SCHED_CPUMASK_DECLARE(allmasks);
7379 struct sched_group **sched_group_nodes = NULL;
7380 int sd_allnodes = 0;
7383 * Allocate the per-node list of sched groups
7385 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7387 if (!sched_group_nodes) {
7388 printk(KERN_WARNING "Can not alloc sched group node list\n");
7393 rd = alloc_rootdomain();
7395 printk(KERN_WARNING "Cannot alloc root domain\n");
7397 kfree(sched_group_nodes);
7402 #if SCHED_CPUMASK_ALLOC
7403 /* get space for all scratch cpumask variables */
7404 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7406 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7409 kfree(sched_group_nodes);
7414 tmpmask = (cpumask_t *)allmasks;
7418 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7422 * Set up domains for cpus specified by the cpu_map.
7424 for_each_cpu_mask_nr(i, *cpu_map) {
7425 struct sched_domain *sd = NULL, *p;
7426 SCHED_CPUMASK_VAR(nodemask, allmasks);
7428 *nodemask = node_to_cpumask(cpu_to_node(i));
7429 cpus_and(*nodemask, *nodemask, *cpu_map);
7432 if (cpus_weight(*cpu_map) >
7433 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7434 sd = &per_cpu(allnodes_domains, i);
7435 SD_INIT(sd, ALLNODES);
7436 set_domain_attribute(sd, attr);
7437 sd->span = *cpu_map;
7438 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7444 sd = &per_cpu(node_domains, i);
7446 set_domain_attribute(sd, attr);
7447 sched_domain_node_span(cpu_to_node(i), &sd->span);
7451 cpus_and(sd->span, sd->span, *cpu_map);
7455 sd = &per_cpu(phys_domains, i);
7457 set_domain_attribute(sd, attr);
7458 sd->span = *nodemask;
7462 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7464 #ifdef CONFIG_SCHED_MC
7466 sd = &per_cpu(core_domains, i);
7468 set_domain_attribute(sd, attr);
7469 sd->span = cpu_coregroup_map(i);
7470 cpus_and(sd->span, sd->span, *cpu_map);
7473 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7476 #ifdef CONFIG_SCHED_SMT
7478 sd = &per_cpu(cpu_domains, i);
7479 SD_INIT(sd, SIBLING);
7480 set_domain_attribute(sd, attr);
7481 sd->span = per_cpu(cpu_sibling_map, i);
7482 cpus_and(sd->span, sd->span, *cpu_map);
7485 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7489 #ifdef CONFIG_SCHED_SMT
7490 /* Set up CPU (sibling) groups */
7491 for_each_cpu_mask_nr(i, *cpu_map) {
7492 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7493 SCHED_CPUMASK_VAR(send_covered, allmasks);
7495 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7496 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7497 if (i != first_cpu(*this_sibling_map))
7500 init_sched_build_groups(this_sibling_map, cpu_map,
7502 send_covered, tmpmask);
7506 #ifdef CONFIG_SCHED_MC
7507 /* Set up multi-core groups */
7508 for_each_cpu_mask_nr(i, *cpu_map) {
7509 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7510 SCHED_CPUMASK_VAR(send_covered, allmasks);
7512 *this_core_map = cpu_coregroup_map(i);
7513 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7514 if (i != first_cpu(*this_core_map))
7517 init_sched_build_groups(this_core_map, cpu_map,
7519 send_covered, tmpmask);
7523 /* Set up physical groups */
7524 for (i = 0; i < nr_node_ids; i++) {
7525 SCHED_CPUMASK_VAR(nodemask, allmasks);
7526 SCHED_CPUMASK_VAR(send_covered, allmasks);
7528 *nodemask = node_to_cpumask(i);
7529 cpus_and(*nodemask, *nodemask, *cpu_map);
7530 if (cpus_empty(*nodemask))
7533 init_sched_build_groups(nodemask, cpu_map,
7535 send_covered, tmpmask);
7539 /* Set up node groups */
7541 SCHED_CPUMASK_VAR(send_covered, allmasks);
7543 init_sched_build_groups(cpu_map, cpu_map,
7544 &cpu_to_allnodes_group,
7545 send_covered, tmpmask);
7548 for (i = 0; i < nr_node_ids; i++) {
7549 /* Set up node groups */
7550 struct sched_group *sg, *prev;
7551 SCHED_CPUMASK_VAR(nodemask, allmasks);
7552 SCHED_CPUMASK_VAR(domainspan, allmasks);
7553 SCHED_CPUMASK_VAR(covered, allmasks);
7556 *nodemask = node_to_cpumask(i);
7557 cpus_clear(*covered);
7559 cpus_and(*nodemask, *nodemask, *cpu_map);
7560 if (cpus_empty(*nodemask)) {
7561 sched_group_nodes[i] = NULL;
7565 sched_domain_node_span(i, domainspan);
7566 cpus_and(*domainspan, *domainspan, *cpu_map);
7568 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7570 printk(KERN_WARNING "Can not alloc domain group for "
7574 sched_group_nodes[i] = sg;
7575 for_each_cpu_mask_nr(j, *nodemask) {
7576 struct sched_domain *sd;
7578 sd = &per_cpu(node_domains, j);
7581 sg->__cpu_power = 0;
7582 sg->cpumask = *nodemask;
7584 cpus_or(*covered, *covered, *nodemask);
7587 for (j = 0; j < nr_node_ids; j++) {
7588 SCHED_CPUMASK_VAR(notcovered, allmasks);
7589 int n = (i + j) % nr_node_ids;
7590 node_to_cpumask_ptr(pnodemask, n);
7592 cpus_complement(*notcovered, *covered);
7593 cpus_and(*tmpmask, *notcovered, *cpu_map);
7594 cpus_and(*tmpmask, *tmpmask, *domainspan);
7595 if (cpus_empty(*tmpmask))
7598 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7599 if (cpus_empty(*tmpmask))
7602 sg = kmalloc_node(sizeof(struct sched_group),
7606 "Can not alloc domain group for node %d\n", j);
7609 sg->__cpu_power = 0;
7610 sg->cpumask = *tmpmask;
7611 sg->next = prev->next;
7612 cpus_or(*covered, *covered, *tmpmask);
7619 /* Calculate CPU power for physical packages and nodes */
7620 #ifdef CONFIG_SCHED_SMT
7621 for_each_cpu_mask_nr(i, *cpu_map) {
7622 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7624 init_sched_groups_power(i, sd);
7627 #ifdef CONFIG_SCHED_MC
7628 for_each_cpu_mask_nr(i, *cpu_map) {
7629 struct sched_domain *sd = &per_cpu(core_domains, i);
7631 init_sched_groups_power(i, sd);
7635 for_each_cpu_mask_nr(i, *cpu_map) {
7636 struct sched_domain *sd = &per_cpu(phys_domains, i);
7638 init_sched_groups_power(i, sd);
7642 for (i = 0; i < nr_node_ids; i++)
7643 init_numa_sched_groups_power(sched_group_nodes[i]);
7646 struct sched_group *sg;
7648 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7650 init_numa_sched_groups_power(sg);
7654 /* Attach the domains */
7655 for_each_cpu_mask_nr(i, *cpu_map) {
7656 struct sched_domain *sd;
7657 #ifdef CONFIG_SCHED_SMT
7658 sd = &per_cpu(cpu_domains, i);
7659 #elif defined(CONFIG_SCHED_MC)
7660 sd = &per_cpu(core_domains, i);
7662 sd = &per_cpu(phys_domains, i);
7664 cpu_attach_domain(sd, rd, i);
7667 SCHED_CPUMASK_FREE((void *)allmasks);
7672 free_sched_groups(cpu_map, tmpmask);
7673 SCHED_CPUMASK_FREE((void *)allmasks);
7678 static int build_sched_domains(const cpumask_t *cpu_map)
7680 return __build_sched_domains(cpu_map, NULL);
7683 static cpumask_t *doms_cur; /* current sched domains */
7684 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7685 static struct sched_domain_attr *dattr_cur;
7686 /* attribues of custom domains in 'doms_cur' */
7689 * Special case: If a kmalloc of a doms_cur partition (array of
7690 * cpumask_t) fails, then fallback to a single sched domain,
7691 * as determined by the single cpumask_t fallback_doms.
7693 static cpumask_t fallback_doms;
7695 void __attribute__((weak)) arch_update_cpu_topology(void)
7700 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7701 * For now this just excludes isolated cpus, but could be used to
7702 * exclude other special cases in the future.
7704 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7708 arch_update_cpu_topology();
7710 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7712 doms_cur = &fallback_doms;
7713 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7715 err = build_sched_domains(doms_cur);
7716 register_sched_domain_sysctl();
7721 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7724 free_sched_groups(cpu_map, tmpmask);
7728 * Detach sched domains from a group of cpus specified in cpu_map
7729 * These cpus will now be attached to the NULL domain
7731 static void detach_destroy_domains(const cpumask_t *cpu_map)
7736 unregister_sched_domain_sysctl();
7738 for_each_cpu_mask_nr(i, *cpu_map)
7739 cpu_attach_domain(NULL, &def_root_domain, i);
7740 synchronize_sched();
7741 arch_destroy_sched_domains(cpu_map, &tmpmask);
7744 /* handle null as "default" */
7745 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7746 struct sched_domain_attr *new, int idx_new)
7748 struct sched_domain_attr tmp;
7755 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7756 new ? (new + idx_new) : &tmp,
7757 sizeof(struct sched_domain_attr));
7761 * Partition sched domains as specified by the 'ndoms_new'
7762 * cpumasks in the array doms_new[] of cpumasks. This compares
7763 * doms_new[] to the current sched domain partitioning, doms_cur[].
7764 * It destroys each deleted domain and builds each new domain.
7766 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7767 * The masks don't intersect (don't overlap.) We should setup one
7768 * sched domain for each mask. CPUs not in any of the cpumasks will
7769 * not be load balanced. If the same cpumask appears both in the
7770 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7773 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7774 * ownership of it and will kfree it when done with it. If the caller
7775 * failed the kmalloc call, then it can pass in doms_new == NULL,
7776 * and partition_sched_domains() will fallback to the single partition
7777 * 'fallback_doms', it also forces the domains to be rebuilt.
7779 * If doms_new==NULL it will be replaced with cpu_online_map.
7780 * ndoms_new==0 is a special case for destroying existing domains.
7781 * It will not create the default domain.
7783 * Call with hotplug lock held
7785 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7786 struct sched_domain_attr *dattr_new)
7790 mutex_lock(&sched_domains_mutex);
7792 /* always unregister in case we don't destroy any domains */
7793 unregister_sched_domain_sysctl();
7795 n = doms_new ? ndoms_new : 0;
7797 /* Destroy deleted domains */
7798 for (i = 0; i < ndoms_cur; i++) {
7799 for (j = 0; j < n; j++) {
7800 if (cpus_equal(doms_cur[i], doms_new[j])
7801 && dattrs_equal(dattr_cur, i, dattr_new, j))
7804 /* no match - a current sched domain not in new doms_new[] */
7805 detach_destroy_domains(doms_cur + i);
7810 if (doms_new == NULL) {
7812 doms_new = &fallback_doms;
7813 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7817 /* Build new domains */
7818 for (i = 0; i < ndoms_new; i++) {
7819 for (j = 0; j < ndoms_cur; j++) {
7820 if (cpus_equal(doms_new[i], doms_cur[j])
7821 && dattrs_equal(dattr_new, i, dattr_cur, j))
7824 /* no match - add a new doms_new */
7825 __build_sched_domains(doms_new + i,
7826 dattr_new ? dattr_new + i : NULL);
7831 /* Remember the new sched domains */
7832 if (doms_cur != &fallback_doms)
7834 kfree(dattr_cur); /* kfree(NULL) is safe */
7835 doms_cur = doms_new;
7836 dattr_cur = dattr_new;
7837 ndoms_cur = ndoms_new;
7839 register_sched_domain_sysctl();
7841 mutex_unlock(&sched_domains_mutex);
7844 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7845 int arch_reinit_sched_domains(void)
7849 /* Destroy domains first to force the rebuild */
7850 partition_sched_domains(0, NULL, NULL);
7852 rebuild_sched_domains();
7858 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7862 if (buf[0] != '0' && buf[0] != '1')
7866 sched_smt_power_savings = (buf[0] == '1');
7868 sched_mc_power_savings = (buf[0] == '1');
7870 ret = arch_reinit_sched_domains();
7872 return ret ? ret : count;
7875 #ifdef CONFIG_SCHED_MC
7876 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7879 return sprintf(page, "%u\n", sched_mc_power_savings);
7881 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7882 const char *buf, size_t count)
7884 return sched_power_savings_store(buf, count, 0);
7886 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7887 sched_mc_power_savings_show,
7888 sched_mc_power_savings_store);
7891 #ifdef CONFIG_SCHED_SMT
7892 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7895 return sprintf(page, "%u\n", sched_smt_power_savings);
7897 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7898 const char *buf, size_t count)
7900 return sched_power_savings_store(buf, count, 1);
7902 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7903 sched_smt_power_savings_show,
7904 sched_smt_power_savings_store);
7907 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7911 #ifdef CONFIG_SCHED_SMT
7913 err = sysfs_create_file(&cls->kset.kobj,
7914 &attr_sched_smt_power_savings.attr);
7916 #ifdef CONFIG_SCHED_MC
7917 if (!err && mc_capable())
7918 err = sysfs_create_file(&cls->kset.kobj,
7919 &attr_sched_mc_power_savings.attr);
7923 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7925 #ifndef CONFIG_CPUSETS
7927 * Add online and remove offline CPUs from the scheduler domains.
7928 * When cpusets are enabled they take over this function.
7930 static int update_sched_domains(struct notifier_block *nfb,
7931 unsigned long action, void *hcpu)
7935 case CPU_ONLINE_FROZEN:
7937 case CPU_DEAD_FROZEN:
7938 partition_sched_domains(1, NULL, NULL);
7947 static int update_runtime(struct notifier_block *nfb,
7948 unsigned long action, void *hcpu)
7950 int cpu = (int)(long)hcpu;
7953 case CPU_DOWN_PREPARE:
7954 case CPU_DOWN_PREPARE_FROZEN:
7955 disable_runtime(cpu_rq(cpu));
7958 case CPU_DOWN_FAILED:
7959 case CPU_DOWN_FAILED_FROZEN:
7961 case CPU_ONLINE_FROZEN:
7962 enable_runtime(cpu_rq(cpu));
7970 void __init sched_init_smp(void)
7972 cpumask_t non_isolated_cpus;
7974 #if defined(CONFIG_NUMA)
7975 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7977 BUG_ON(sched_group_nodes_bycpu == NULL);
7980 mutex_lock(&sched_domains_mutex);
7981 arch_init_sched_domains(&cpu_online_map);
7982 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7983 if (cpus_empty(non_isolated_cpus))
7984 cpu_set(smp_processor_id(), non_isolated_cpus);
7985 mutex_unlock(&sched_domains_mutex);
7988 #ifndef CONFIG_CPUSETS
7989 /* XXX: Theoretical race here - CPU may be hotplugged now */
7990 hotcpu_notifier(update_sched_domains, 0);
7993 /* RT runtime code needs to handle some hotplug events */
7994 hotcpu_notifier(update_runtime, 0);
7998 /* Move init over to a non-isolated CPU */
7999 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
8001 sched_init_granularity();
8004 void __init sched_init_smp(void)
8006 sched_init_granularity();
8008 #endif /* CONFIG_SMP */
8010 int in_sched_functions(unsigned long addr)
8012 return in_lock_functions(addr) ||
8013 (addr >= (unsigned long)__sched_text_start
8014 && addr < (unsigned long)__sched_text_end);
8017 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8019 cfs_rq->tasks_timeline = RB_ROOT;
8020 INIT_LIST_HEAD(&cfs_rq->tasks);
8021 #ifdef CONFIG_FAIR_GROUP_SCHED
8024 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8027 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8029 struct rt_prio_array *array;
8032 array = &rt_rq->active;
8033 for (i = 0; i < MAX_RT_PRIO; i++) {
8034 INIT_LIST_HEAD(array->queue + i);
8035 __clear_bit(i, array->bitmap);
8037 /* delimiter for bitsearch: */
8038 __set_bit(MAX_RT_PRIO, array->bitmap);
8040 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8041 rt_rq->highest_prio = MAX_RT_PRIO;
8044 rt_rq->rt_nr_migratory = 0;
8045 rt_rq->overloaded = 0;
8049 rt_rq->rt_throttled = 0;
8050 rt_rq->rt_runtime = 0;
8051 spin_lock_init(&rt_rq->rt_runtime_lock);
8053 #ifdef CONFIG_RT_GROUP_SCHED
8054 rt_rq->rt_nr_boosted = 0;
8059 #ifdef CONFIG_FAIR_GROUP_SCHED
8060 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8061 struct sched_entity *se, int cpu, int add,
8062 struct sched_entity *parent)
8064 struct rq *rq = cpu_rq(cpu);
8065 tg->cfs_rq[cpu] = cfs_rq;
8066 init_cfs_rq(cfs_rq, rq);
8069 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8072 /* se could be NULL for init_task_group */
8077 se->cfs_rq = &rq->cfs;
8079 se->cfs_rq = parent->my_q;
8082 se->load.weight = tg->shares;
8083 se->load.inv_weight = 0;
8084 se->parent = parent;
8088 #ifdef CONFIG_RT_GROUP_SCHED
8089 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8090 struct sched_rt_entity *rt_se, int cpu, int add,
8091 struct sched_rt_entity *parent)
8093 struct rq *rq = cpu_rq(cpu);
8095 tg->rt_rq[cpu] = rt_rq;
8096 init_rt_rq(rt_rq, rq);
8098 rt_rq->rt_se = rt_se;
8099 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8101 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8103 tg->rt_se[cpu] = rt_se;
8108 rt_se->rt_rq = &rq->rt;
8110 rt_se->rt_rq = parent->my_q;
8112 rt_se->my_q = rt_rq;
8113 rt_se->parent = parent;
8114 INIT_LIST_HEAD(&rt_se->run_list);
8118 void __init sched_init(void)
8121 unsigned long alloc_size = 0, ptr;
8123 #ifdef CONFIG_FAIR_GROUP_SCHED
8124 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8126 #ifdef CONFIG_RT_GROUP_SCHED
8127 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8129 #ifdef CONFIG_USER_SCHED
8133 * As sched_init() is called before page_alloc is setup,
8134 * we use alloc_bootmem().
8137 ptr = (unsigned long)alloc_bootmem(alloc_size);
8139 #ifdef CONFIG_FAIR_GROUP_SCHED
8140 init_task_group.se = (struct sched_entity **)ptr;
8141 ptr += nr_cpu_ids * sizeof(void **);
8143 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8144 ptr += nr_cpu_ids * sizeof(void **);
8146 #ifdef CONFIG_USER_SCHED
8147 root_task_group.se = (struct sched_entity **)ptr;
8148 ptr += nr_cpu_ids * sizeof(void **);
8150 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8151 ptr += nr_cpu_ids * sizeof(void **);
8152 #endif /* CONFIG_USER_SCHED */
8153 #endif /* CONFIG_FAIR_GROUP_SCHED */
8154 #ifdef CONFIG_RT_GROUP_SCHED
8155 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8156 ptr += nr_cpu_ids * sizeof(void **);
8158 init_task_group.rt_rq = (struct rt_rq **)ptr;
8159 ptr += nr_cpu_ids * sizeof(void **);
8161 #ifdef CONFIG_USER_SCHED
8162 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8163 ptr += nr_cpu_ids * sizeof(void **);
8165 root_task_group.rt_rq = (struct rt_rq **)ptr;
8166 ptr += nr_cpu_ids * sizeof(void **);
8167 #endif /* CONFIG_USER_SCHED */
8168 #endif /* CONFIG_RT_GROUP_SCHED */
8172 init_defrootdomain();
8175 init_rt_bandwidth(&def_rt_bandwidth,
8176 global_rt_period(), global_rt_runtime());
8178 #ifdef CONFIG_RT_GROUP_SCHED
8179 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8180 global_rt_period(), global_rt_runtime());
8181 #ifdef CONFIG_USER_SCHED
8182 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8183 global_rt_period(), RUNTIME_INF);
8184 #endif /* CONFIG_USER_SCHED */
8185 #endif /* CONFIG_RT_GROUP_SCHED */
8187 #ifdef CONFIG_GROUP_SCHED
8188 list_add(&init_task_group.list, &task_groups);
8189 INIT_LIST_HEAD(&init_task_group.children);
8191 #ifdef CONFIG_USER_SCHED
8192 INIT_LIST_HEAD(&root_task_group.children);
8193 init_task_group.parent = &root_task_group;
8194 list_add(&init_task_group.siblings, &root_task_group.children);
8195 #endif /* CONFIG_USER_SCHED */
8196 #endif /* CONFIG_GROUP_SCHED */
8198 for_each_possible_cpu(i) {
8202 spin_lock_init(&rq->lock);
8204 init_cfs_rq(&rq->cfs, rq);
8205 init_rt_rq(&rq->rt, rq);
8206 #ifdef CONFIG_FAIR_GROUP_SCHED
8207 init_task_group.shares = init_task_group_load;
8208 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8209 #ifdef CONFIG_CGROUP_SCHED
8211 * How much cpu bandwidth does init_task_group get?
8213 * In case of task-groups formed thr' the cgroup filesystem, it
8214 * gets 100% of the cpu resources in the system. This overall
8215 * system cpu resource is divided among the tasks of
8216 * init_task_group and its child task-groups in a fair manner,
8217 * based on each entity's (task or task-group's) weight
8218 * (se->load.weight).
8220 * In other words, if init_task_group has 10 tasks of weight
8221 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8222 * then A0's share of the cpu resource is:
8224 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8226 * We achieve this by letting init_task_group's tasks sit
8227 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8229 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8230 #elif defined CONFIG_USER_SCHED
8231 root_task_group.shares = NICE_0_LOAD;
8232 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8234 * In case of task-groups formed thr' the user id of tasks,
8235 * init_task_group represents tasks belonging to root user.
8236 * Hence it forms a sibling of all subsequent groups formed.
8237 * In this case, init_task_group gets only a fraction of overall
8238 * system cpu resource, based on the weight assigned to root
8239 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8240 * by letting tasks of init_task_group sit in a separate cfs_rq
8241 * (init_cfs_rq) and having one entity represent this group of
8242 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8244 init_tg_cfs_entry(&init_task_group,
8245 &per_cpu(init_cfs_rq, i),
8246 &per_cpu(init_sched_entity, i), i, 1,
8247 root_task_group.se[i]);
8250 #endif /* CONFIG_FAIR_GROUP_SCHED */
8252 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8253 #ifdef CONFIG_RT_GROUP_SCHED
8254 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8255 #ifdef CONFIG_CGROUP_SCHED
8256 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8257 #elif defined CONFIG_USER_SCHED
8258 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8259 init_tg_rt_entry(&init_task_group,
8260 &per_cpu(init_rt_rq, i),
8261 &per_cpu(init_sched_rt_entity, i), i, 1,
8262 root_task_group.rt_se[i]);
8266 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8267 rq->cpu_load[j] = 0;
8271 rq->active_balance = 0;
8272 rq->next_balance = jiffies;
8276 rq->migration_thread = NULL;
8277 INIT_LIST_HEAD(&rq->migration_queue);
8278 rq_attach_root(rq, &def_root_domain);
8281 atomic_set(&rq->nr_iowait, 0);
8284 set_load_weight(&init_task);
8286 #ifdef CONFIG_PREEMPT_NOTIFIERS
8287 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8291 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8294 #ifdef CONFIG_RT_MUTEXES
8295 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8299 * The boot idle thread does lazy MMU switching as well:
8301 atomic_inc(&init_mm.mm_count);
8302 enter_lazy_tlb(&init_mm, current);
8305 * Make us the idle thread. Technically, schedule() should not be
8306 * called from this thread, however somewhere below it might be,
8307 * but because we are the idle thread, we just pick up running again
8308 * when this runqueue becomes "idle".
8310 init_idle(current, smp_processor_id());
8312 * During early bootup we pretend to be a normal task:
8314 current->sched_class = &fair_sched_class;
8316 scheduler_running = 1;
8319 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8320 void __might_sleep(char *file, int line)
8323 static unsigned long prev_jiffy; /* ratelimiting */
8325 if ((!in_atomic() && !irqs_disabled()) ||
8326 system_state != SYSTEM_RUNNING || oops_in_progress)
8328 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8330 prev_jiffy = jiffies;
8333 "BUG: sleeping function called from invalid context at %s:%d\n",
8336 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8337 in_atomic(), irqs_disabled(),
8338 current->pid, current->comm);
8340 debug_show_held_locks(current);
8341 if (irqs_disabled())
8342 print_irqtrace_events(current);
8346 EXPORT_SYMBOL(__might_sleep);
8349 #ifdef CONFIG_MAGIC_SYSRQ
8350 static void normalize_task(struct rq *rq, struct task_struct *p)
8354 update_rq_clock(rq);
8355 on_rq = p->se.on_rq;
8357 deactivate_task(rq, p, 0);
8358 __setscheduler(rq, p, SCHED_NORMAL, 0);
8360 activate_task(rq, p, 0);
8361 resched_task(rq->curr);
8365 void normalize_rt_tasks(void)
8367 struct task_struct *g, *p;
8368 unsigned long flags;
8371 read_lock_irqsave(&tasklist_lock, flags);
8372 do_each_thread(g, p) {
8374 * Only normalize user tasks:
8379 p->se.exec_start = 0;
8380 #ifdef CONFIG_SCHEDSTATS
8381 p->se.wait_start = 0;
8382 p->se.sleep_start = 0;
8383 p->se.block_start = 0;
8388 * Renice negative nice level userspace
8391 if (TASK_NICE(p) < 0 && p->mm)
8392 set_user_nice(p, 0);
8396 spin_lock(&p->pi_lock);
8397 rq = __task_rq_lock(p);
8399 normalize_task(rq, p);
8401 __task_rq_unlock(rq);
8402 spin_unlock(&p->pi_lock);
8403 } while_each_thread(g, p);
8405 read_unlock_irqrestore(&tasklist_lock, flags);
8408 #endif /* CONFIG_MAGIC_SYSRQ */
8412 * These functions are only useful for the IA64 MCA handling.
8414 * They can only be called when the whole system has been
8415 * stopped - every CPU needs to be quiescent, and no scheduling
8416 * activity can take place. Using them for anything else would
8417 * be a serious bug, and as a result, they aren't even visible
8418 * under any other configuration.
8422 * curr_task - return the current task for a given cpu.
8423 * @cpu: the processor in question.
8425 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8427 struct task_struct *curr_task(int cpu)
8429 return cpu_curr(cpu);
8433 * set_curr_task - set the current task for a given cpu.
8434 * @cpu: the processor in question.
8435 * @p: the task pointer to set.
8437 * Description: This function must only be used when non-maskable interrupts
8438 * are serviced on a separate stack. It allows the architecture to switch the
8439 * notion of the current task on a cpu in a non-blocking manner. This function
8440 * must be called with all CPU's synchronized, and interrupts disabled, the
8441 * and caller must save the original value of the current task (see
8442 * curr_task() above) and restore that value before reenabling interrupts and
8443 * re-starting the system.
8445 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8447 void set_curr_task(int cpu, struct task_struct *p)
8454 #ifdef CONFIG_FAIR_GROUP_SCHED
8455 static void free_fair_sched_group(struct task_group *tg)
8459 for_each_possible_cpu(i) {
8461 kfree(tg->cfs_rq[i]);
8471 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8473 struct cfs_rq *cfs_rq;
8474 struct sched_entity *se, *parent_se;
8478 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8481 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8485 tg->shares = NICE_0_LOAD;
8487 for_each_possible_cpu(i) {
8490 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8491 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8495 se = kmalloc_node(sizeof(struct sched_entity),
8496 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8500 parent_se = parent ? parent->se[i] : NULL;
8501 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8510 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8512 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8513 &cpu_rq(cpu)->leaf_cfs_rq_list);
8516 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8518 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8520 #else /* !CONFG_FAIR_GROUP_SCHED */
8521 static inline void free_fair_sched_group(struct task_group *tg)
8526 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8531 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8535 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8538 #endif /* CONFIG_FAIR_GROUP_SCHED */
8540 #ifdef CONFIG_RT_GROUP_SCHED
8541 static void free_rt_sched_group(struct task_group *tg)
8545 destroy_rt_bandwidth(&tg->rt_bandwidth);
8547 for_each_possible_cpu(i) {
8549 kfree(tg->rt_rq[i]);
8551 kfree(tg->rt_se[i]);
8559 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8561 struct rt_rq *rt_rq;
8562 struct sched_rt_entity *rt_se, *parent_se;
8566 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8569 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8573 init_rt_bandwidth(&tg->rt_bandwidth,
8574 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8576 for_each_possible_cpu(i) {
8579 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8580 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8584 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8585 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8589 parent_se = parent ? parent->rt_se[i] : NULL;
8590 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8599 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8601 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8602 &cpu_rq(cpu)->leaf_rt_rq_list);
8605 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8607 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8609 #else /* !CONFIG_RT_GROUP_SCHED */
8610 static inline void free_rt_sched_group(struct task_group *tg)
8615 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8620 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8624 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8627 #endif /* CONFIG_RT_GROUP_SCHED */
8629 #ifdef CONFIG_GROUP_SCHED
8630 static void free_sched_group(struct task_group *tg)
8632 free_fair_sched_group(tg);
8633 free_rt_sched_group(tg);
8637 /* allocate runqueue etc for a new task group */
8638 struct task_group *sched_create_group(struct task_group *parent)
8640 struct task_group *tg;
8641 unsigned long flags;
8644 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8646 return ERR_PTR(-ENOMEM);
8648 if (!alloc_fair_sched_group(tg, parent))
8651 if (!alloc_rt_sched_group(tg, parent))
8654 spin_lock_irqsave(&task_group_lock, flags);
8655 for_each_possible_cpu(i) {
8656 register_fair_sched_group(tg, i);
8657 register_rt_sched_group(tg, i);
8659 list_add_rcu(&tg->list, &task_groups);
8661 WARN_ON(!parent); /* root should already exist */
8663 tg->parent = parent;
8664 INIT_LIST_HEAD(&tg->children);
8665 list_add_rcu(&tg->siblings, &parent->children);
8666 spin_unlock_irqrestore(&task_group_lock, flags);
8671 free_sched_group(tg);
8672 return ERR_PTR(-ENOMEM);
8675 /* rcu callback to free various structures associated with a task group */
8676 static void free_sched_group_rcu(struct rcu_head *rhp)
8678 /* now it should be safe to free those cfs_rqs */
8679 free_sched_group(container_of(rhp, struct task_group, rcu));
8682 /* Destroy runqueue etc associated with a task group */
8683 void sched_destroy_group(struct task_group *tg)
8685 unsigned long flags;
8688 spin_lock_irqsave(&task_group_lock, flags);
8689 for_each_possible_cpu(i) {
8690 unregister_fair_sched_group(tg, i);
8691 unregister_rt_sched_group(tg, i);
8693 list_del_rcu(&tg->list);
8694 list_del_rcu(&tg->siblings);
8695 spin_unlock_irqrestore(&task_group_lock, flags);
8697 /* wait for possible concurrent references to cfs_rqs complete */
8698 call_rcu(&tg->rcu, free_sched_group_rcu);
8701 /* change task's runqueue when it moves between groups.
8702 * The caller of this function should have put the task in its new group
8703 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8704 * reflect its new group.
8706 void sched_move_task(struct task_struct *tsk)
8709 unsigned long flags;
8712 rq = task_rq_lock(tsk, &flags);
8714 update_rq_clock(rq);
8716 running = task_current(rq, tsk);
8717 on_rq = tsk->se.on_rq;
8720 dequeue_task(rq, tsk, 0);
8721 if (unlikely(running))
8722 tsk->sched_class->put_prev_task(rq, tsk);
8724 set_task_rq(tsk, task_cpu(tsk));
8726 #ifdef CONFIG_FAIR_GROUP_SCHED
8727 if (tsk->sched_class->moved_group)
8728 tsk->sched_class->moved_group(tsk);
8731 if (unlikely(running))
8732 tsk->sched_class->set_curr_task(rq);
8734 enqueue_task(rq, tsk, 0);
8736 task_rq_unlock(rq, &flags);
8738 #endif /* CONFIG_GROUP_SCHED */
8740 #ifdef CONFIG_FAIR_GROUP_SCHED
8741 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8743 struct cfs_rq *cfs_rq = se->cfs_rq;
8748 dequeue_entity(cfs_rq, se, 0);
8750 se->load.weight = shares;
8751 se->load.inv_weight = 0;
8754 enqueue_entity(cfs_rq, se, 0);
8757 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8759 struct cfs_rq *cfs_rq = se->cfs_rq;
8760 struct rq *rq = cfs_rq->rq;
8761 unsigned long flags;
8763 spin_lock_irqsave(&rq->lock, flags);
8764 __set_se_shares(se, shares);
8765 spin_unlock_irqrestore(&rq->lock, flags);
8768 static DEFINE_MUTEX(shares_mutex);
8770 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8773 unsigned long flags;
8776 * We can't change the weight of the root cgroup.
8781 if (shares < MIN_SHARES)
8782 shares = MIN_SHARES;
8783 else if (shares > MAX_SHARES)
8784 shares = MAX_SHARES;
8786 mutex_lock(&shares_mutex);
8787 if (tg->shares == shares)
8790 spin_lock_irqsave(&task_group_lock, flags);
8791 for_each_possible_cpu(i)
8792 unregister_fair_sched_group(tg, i);
8793 list_del_rcu(&tg->siblings);
8794 spin_unlock_irqrestore(&task_group_lock, flags);
8796 /* wait for any ongoing reference to this group to finish */
8797 synchronize_sched();
8800 * Now we are free to modify the group's share on each cpu
8801 * w/o tripping rebalance_share or load_balance_fair.
8803 tg->shares = shares;
8804 for_each_possible_cpu(i) {
8808 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8809 set_se_shares(tg->se[i], shares);
8813 * Enable load balance activity on this group, by inserting it back on
8814 * each cpu's rq->leaf_cfs_rq_list.
8816 spin_lock_irqsave(&task_group_lock, flags);
8817 for_each_possible_cpu(i)
8818 register_fair_sched_group(tg, i);
8819 list_add_rcu(&tg->siblings, &tg->parent->children);
8820 spin_unlock_irqrestore(&task_group_lock, flags);
8822 mutex_unlock(&shares_mutex);
8826 unsigned long sched_group_shares(struct task_group *tg)
8832 #ifdef CONFIG_RT_GROUP_SCHED
8834 * Ensure that the real time constraints are schedulable.
8836 static DEFINE_MUTEX(rt_constraints_mutex);
8838 static unsigned long to_ratio(u64 period, u64 runtime)
8840 if (runtime == RUNTIME_INF)
8843 return div64_u64(runtime << 20, period);
8846 /* Must be called with tasklist_lock held */
8847 static inline int tg_has_rt_tasks(struct task_group *tg)
8849 struct task_struct *g, *p;
8851 do_each_thread(g, p) {
8852 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8854 } while_each_thread(g, p);
8859 struct rt_schedulable_data {
8860 struct task_group *tg;
8865 static int tg_schedulable(struct task_group *tg, void *data)
8867 struct rt_schedulable_data *d = data;
8868 struct task_group *child;
8869 unsigned long total, sum = 0;
8870 u64 period, runtime;
8872 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8873 runtime = tg->rt_bandwidth.rt_runtime;
8876 period = d->rt_period;
8877 runtime = d->rt_runtime;
8881 * Cannot have more runtime than the period.
8883 if (runtime > period && runtime != RUNTIME_INF)
8887 * Ensure we don't starve existing RT tasks.
8889 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8892 total = to_ratio(period, runtime);
8895 * Nobody can have more than the global setting allows.
8897 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8901 * The sum of our children's runtime should not exceed our own.
8903 list_for_each_entry_rcu(child, &tg->children, siblings) {
8904 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8905 runtime = child->rt_bandwidth.rt_runtime;
8907 if (child == d->tg) {
8908 period = d->rt_period;
8909 runtime = d->rt_runtime;
8912 sum += to_ratio(period, runtime);
8921 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8923 struct rt_schedulable_data data = {
8925 .rt_period = period,
8926 .rt_runtime = runtime,
8929 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8932 static int tg_set_bandwidth(struct task_group *tg,
8933 u64 rt_period, u64 rt_runtime)
8937 mutex_lock(&rt_constraints_mutex);
8938 read_lock(&tasklist_lock);
8939 err = __rt_schedulable(tg, rt_period, rt_runtime);
8943 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8944 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8945 tg->rt_bandwidth.rt_runtime = rt_runtime;
8947 for_each_possible_cpu(i) {
8948 struct rt_rq *rt_rq = tg->rt_rq[i];
8950 spin_lock(&rt_rq->rt_runtime_lock);
8951 rt_rq->rt_runtime = rt_runtime;
8952 spin_unlock(&rt_rq->rt_runtime_lock);
8954 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8956 read_unlock(&tasklist_lock);
8957 mutex_unlock(&rt_constraints_mutex);
8962 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8964 u64 rt_runtime, rt_period;
8966 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8967 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8968 if (rt_runtime_us < 0)
8969 rt_runtime = RUNTIME_INF;
8971 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8974 long sched_group_rt_runtime(struct task_group *tg)
8978 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8981 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8982 do_div(rt_runtime_us, NSEC_PER_USEC);
8983 return rt_runtime_us;
8986 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8988 u64 rt_runtime, rt_period;
8990 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8991 rt_runtime = tg->rt_bandwidth.rt_runtime;
8996 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8999 long sched_group_rt_period(struct task_group *tg)
9003 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9004 do_div(rt_period_us, NSEC_PER_USEC);
9005 return rt_period_us;
9008 static int sched_rt_global_constraints(void)
9010 u64 runtime, period;
9013 if (sysctl_sched_rt_period <= 0)
9016 runtime = global_rt_runtime();
9017 period = global_rt_period();
9020 * Sanity check on the sysctl variables.
9022 if (runtime > period && runtime != RUNTIME_INF)
9025 mutex_lock(&rt_constraints_mutex);
9026 read_lock(&tasklist_lock);
9027 ret = __rt_schedulable(NULL, 0, 0);
9028 read_unlock(&tasklist_lock);
9029 mutex_unlock(&rt_constraints_mutex);
9033 #else /* !CONFIG_RT_GROUP_SCHED */
9034 static int sched_rt_global_constraints(void)
9036 unsigned long flags;
9039 if (sysctl_sched_rt_period <= 0)
9042 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9043 for_each_possible_cpu(i) {
9044 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9046 spin_lock(&rt_rq->rt_runtime_lock);
9047 rt_rq->rt_runtime = global_rt_runtime();
9048 spin_unlock(&rt_rq->rt_runtime_lock);
9050 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9054 #endif /* CONFIG_RT_GROUP_SCHED */
9056 int sched_rt_handler(struct ctl_table *table, int write,
9057 struct file *filp, void __user *buffer, size_t *lenp,
9061 int old_period, old_runtime;
9062 static DEFINE_MUTEX(mutex);
9065 old_period = sysctl_sched_rt_period;
9066 old_runtime = sysctl_sched_rt_runtime;
9068 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9070 if (!ret && write) {
9071 ret = sched_rt_global_constraints();
9073 sysctl_sched_rt_period = old_period;
9074 sysctl_sched_rt_runtime = old_runtime;
9076 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9077 def_rt_bandwidth.rt_period =
9078 ns_to_ktime(global_rt_period());
9081 mutex_unlock(&mutex);
9086 #ifdef CONFIG_CGROUP_SCHED
9088 /* return corresponding task_group object of a cgroup */
9089 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9091 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9092 struct task_group, css);
9095 static struct cgroup_subsys_state *
9096 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9098 struct task_group *tg, *parent;
9100 if (!cgrp->parent) {
9101 /* This is early initialization for the top cgroup */
9102 return &init_task_group.css;
9105 parent = cgroup_tg(cgrp->parent);
9106 tg = sched_create_group(parent);
9108 return ERR_PTR(-ENOMEM);
9114 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9116 struct task_group *tg = cgroup_tg(cgrp);
9118 sched_destroy_group(tg);
9122 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9123 struct task_struct *tsk)
9125 #ifdef CONFIG_RT_GROUP_SCHED
9126 /* Don't accept realtime tasks when there is no way for them to run */
9127 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9130 /* We don't support RT-tasks being in separate groups */
9131 if (tsk->sched_class != &fair_sched_class)
9139 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9140 struct cgroup *old_cont, struct task_struct *tsk)
9142 sched_move_task(tsk);
9145 #ifdef CONFIG_FAIR_GROUP_SCHED
9146 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9149 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9152 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9154 struct task_group *tg = cgroup_tg(cgrp);
9156 return (u64) tg->shares;
9158 #endif /* CONFIG_FAIR_GROUP_SCHED */
9160 #ifdef CONFIG_RT_GROUP_SCHED
9161 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9164 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9167 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9169 return sched_group_rt_runtime(cgroup_tg(cgrp));
9172 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9175 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9178 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9180 return sched_group_rt_period(cgroup_tg(cgrp));
9182 #endif /* CONFIG_RT_GROUP_SCHED */
9184 static struct cftype cpu_files[] = {
9185 #ifdef CONFIG_FAIR_GROUP_SCHED
9188 .read_u64 = cpu_shares_read_u64,
9189 .write_u64 = cpu_shares_write_u64,
9192 #ifdef CONFIG_RT_GROUP_SCHED
9194 .name = "rt_runtime_us",
9195 .read_s64 = cpu_rt_runtime_read,
9196 .write_s64 = cpu_rt_runtime_write,
9199 .name = "rt_period_us",
9200 .read_u64 = cpu_rt_period_read_uint,
9201 .write_u64 = cpu_rt_period_write_uint,
9206 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9208 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9211 struct cgroup_subsys cpu_cgroup_subsys = {
9213 .create = cpu_cgroup_create,
9214 .destroy = cpu_cgroup_destroy,
9215 .can_attach = cpu_cgroup_can_attach,
9216 .attach = cpu_cgroup_attach,
9217 .populate = cpu_cgroup_populate,
9218 .subsys_id = cpu_cgroup_subsys_id,
9222 #endif /* CONFIG_CGROUP_SCHED */
9224 #ifdef CONFIG_CGROUP_CPUACCT
9227 * CPU accounting code for task groups.
9229 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9230 * (balbir@in.ibm.com).
9233 /* track cpu usage of a group of tasks */
9235 struct cgroup_subsys_state css;
9236 /* cpuusage holds pointer to a u64-type object on every cpu */
9240 struct cgroup_subsys cpuacct_subsys;
9242 /* return cpu accounting group corresponding to this container */
9243 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9245 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9246 struct cpuacct, css);
9249 /* return cpu accounting group to which this task belongs */
9250 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9252 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9253 struct cpuacct, css);
9256 /* create a new cpu accounting group */
9257 static struct cgroup_subsys_state *cpuacct_create(
9258 struct cgroup_subsys *ss, struct cgroup *cgrp)
9260 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9263 return ERR_PTR(-ENOMEM);
9265 ca->cpuusage = alloc_percpu(u64);
9266 if (!ca->cpuusage) {
9268 return ERR_PTR(-ENOMEM);
9274 /* destroy an existing cpu accounting group */
9276 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9278 struct cpuacct *ca = cgroup_ca(cgrp);
9280 free_percpu(ca->cpuusage);
9284 /* return total cpu usage (in nanoseconds) of a group */
9285 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9287 struct cpuacct *ca = cgroup_ca(cgrp);
9288 u64 totalcpuusage = 0;
9291 for_each_possible_cpu(i) {
9292 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9295 * Take rq->lock to make 64-bit addition safe on 32-bit
9298 spin_lock_irq(&cpu_rq(i)->lock);
9299 totalcpuusage += *cpuusage;
9300 spin_unlock_irq(&cpu_rq(i)->lock);
9303 return totalcpuusage;
9306 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9309 struct cpuacct *ca = cgroup_ca(cgrp);
9318 for_each_possible_cpu(i) {
9319 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9321 spin_lock_irq(&cpu_rq(i)->lock);
9323 spin_unlock_irq(&cpu_rq(i)->lock);
9329 static struct cftype files[] = {
9332 .read_u64 = cpuusage_read,
9333 .write_u64 = cpuusage_write,
9337 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9339 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9343 * charge this task's execution time to its accounting group.
9345 * called with rq->lock held.
9347 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9351 if (!cpuacct_subsys.active)
9356 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9358 *cpuusage += cputime;
9362 struct cgroup_subsys cpuacct_subsys = {
9364 .create = cpuacct_create,
9365 .destroy = cpuacct_destroy,
9366 .populate = cpuacct_populate,
9367 .subsys_id = cpuacct_subsys_id,
9369 #endif /* CONFIG_CGROUP_CPUACCT */