4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/bootmem.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
75 #include <asm/irq_regs.h>
78 * Convert user-nice values [ -20 ... 0 ... 19 ]
79 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
82 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
83 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
84 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
87 * 'User priority' is the nice value converted to something we
88 * can work with better when scaling various scheduler parameters,
89 * it's a [ 0 ... 39 ] range.
91 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
92 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
93 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
96 * Helpers for converting nanosecond timing to jiffy resolution
98 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
100 #define NICE_0_LOAD SCHED_LOAD_SCALE
101 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
104 * These are the 'tuning knobs' of the scheduler:
106 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
107 * Timeslices get refilled after they expire.
109 #define DEF_TIMESLICE (100 * HZ / 1000)
112 * single value that denotes runtime == period, ie unlimited time.
114 #define RUNTIME_INF ((u64)~0ULL)
118 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
119 * Since cpu_power is a 'constant', we can use a reciprocal divide.
121 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
123 return reciprocal_divide(load, sg->reciprocal_cpu_power);
127 * Each time a sched group cpu_power is changed,
128 * we must compute its reciprocal value
130 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
132 sg->__cpu_power += val;
133 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
137 static inline int rt_policy(int policy)
139 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
144 static inline int task_has_rt_policy(struct task_struct *p)
146 return rt_policy(p->policy);
150 * This is the priority-queue data structure of the RT scheduling class:
152 struct rt_prio_array {
153 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
154 struct list_head queue[MAX_RT_PRIO];
157 struct rt_bandwidth {
158 /* nests inside the rq lock: */
159 spinlock_t rt_runtime_lock;
162 struct hrtimer rt_period_timer;
165 static struct rt_bandwidth def_rt_bandwidth;
167 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
169 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
171 struct rt_bandwidth *rt_b =
172 container_of(timer, struct rt_bandwidth, rt_period_timer);
178 now = hrtimer_cb_get_time(timer);
179 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
184 idle = do_sched_rt_period_timer(rt_b, overrun);
187 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
191 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
193 rt_b->rt_period = ns_to_ktime(period);
194 rt_b->rt_runtime = runtime;
196 spin_lock_init(&rt_b->rt_runtime_lock);
198 hrtimer_init(&rt_b->rt_period_timer,
199 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
200 rt_b->rt_period_timer.function = sched_rt_period_timer;
201 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
204 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
208 if (rt_b->rt_runtime == RUNTIME_INF)
211 if (hrtimer_active(&rt_b->rt_period_timer))
214 spin_lock(&rt_b->rt_runtime_lock);
216 if (hrtimer_active(&rt_b->rt_period_timer))
219 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
220 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
221 hrtimer_start(&rt_b->rt_period_timer,
222 rt_b->rt_period_timer.expires,
225 spin_unlock(&rt_b->rt_runtime_lock);
228 #ifdef CONFIG_RT_GROUP_SCHED
229 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
231 hrtimer_cancel(&rt_b->rt_period_timer);
236 * sched_domains_mutex serializes calls to arch_init_sched_domains,
237 * detach_destroy_domains and partition_sched_domains.
239 static DEFINE_MUTEX(sched_domains_mutex);
241 #ifdef CONFIG_GROUP_SCHED
243 #include <linux/cgroup.h>
247 static LIST_HEAD(task_groups);
249 /* task group related information */
251 #ifdef CONFIG_CGROUP_SCHED
252 struct cgroup_subsys_state css;
255 #ifdef CONFIG_FAIR_GROUP_SCHED
256 /* schedulable entities of this group on each cpu */
257 struct sched_entity **se;
258 /* runqueue "owned" by this group on each cpu */
259 struct cfs_rq **cfs_rq;
260 unsigned long shares;
263 #ifdef CONFIG_RT_GROUP_SCHED
264 struct sched_rt_entity **rt_se;
265 struct rt_rq **rt_rq;
267 struct rt_bandwidth rt_bandwidth;
271 struct list_head list;
273 struct task_group *parent;
274 struct list_head siblings;
275 struct list_head children;
278 #ifdef CONFIG_USER_SCHED
282 * Every UID task group (including init_task_group aka UID-0) will
283 * be a child to this group.
285 struct task_group root_task_group;
287 #ifdef CONFIG_FAIR_GROUP_SCHED
288 /* Default task group's sched entity on each cpu */
289 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
290 /* Default task group's cfs_rq on each cpu */
291 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
294 #ifdef CONFIG_RT_GROUP_SCHED
295 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
296 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
299 #define root_task_group init_task_group
302 /* task_group_lock serializes add/remove of task groups and also changes to
303 * a task group's cpu shares.
305 static DEFINE_SPINLOCK(task_group_lock);
307 #ifdef CONFIG_FAIR_GROUP_SCHED
308 #ifdef CONFIG_USER_SCHED
309 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
311 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
315 * A weight of 0, 1 or ULONG_MAX can cause arithmetics problems.
316 * (The default weight is 1024 - so there's no practical
317 * limitation from this.)
320 #define MAX_SHARES (ULONG_MAX - 1)
322 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
325 /* Default task group.
326 * Every task in system belong to this group at bootup.
328 struct task_group init_task_group;
330 /* return group to which a task belongs */
331 static inline struct task_group *task_group(struct task_struct *p)
333 struct task_group *tg;
335 #ifdef CONFIG_USER_SCHED
337 #elif defined(CONFIG_CGROUP_SCHED)
338 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
339 struct task_group, css);
341 tg = &init_task_group;
346 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
347 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
349 #ifdef CONFIG_FAIR_GROUP_SCHED
350 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
351 p->se.parent = task_group(p)->se[cpu];
354 #ifdef CONFIG_RT_GROUP_SCHED
355 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
356 p->rt.parent = task_group(p)->rt_se[cpu];
362 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
364 #endif /* CONFIG_GROUP_SCHED */
366 /* CFS-related fields in a runqueue */
368 struct load_weight load;
369 unsigned long nr_running;
374 struct rb_root tasks_timeline;
375 struct rb_node *rb_leftmost;
377 struct list_head tasks;
378 struct list_head *balance_iterator;
381 * 'curr' points to currently running entity on this cfs_rq.
382 * It is set to NULL otherwise (i.e when none are currently running).
384 struct sched_entity *curr, *next;
386 unsigned long nr_spread_over;
388 #ifdef CONFIG_FAIR_GROUP_SCHED
389 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
392 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
393 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
394 * (like users, containers etc.)
396 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
397 * list is used during load balance.
399 struct list_head leaf_cfs_rq_list;
400 struct task_group *tg; /* group that "owns" this runqueue */
403 unsigned long task_weight;
404 unsigned long shares;
406 * We need space to build a sched_domain wide view of the full task
407 * group tree, in order to avoid depending on dynamic memory allocation
408 * during the load balancing we place this in the per cpu task group
409 * hierarchy. This limits the load balancing to one instance per cpu,
410 * but more should not be needed anyway.
412 struct aggregate_struct {
414 * load = weight(cpus) * f(tg)
416 * Where f(tg) is the recursive weight fraction assigned to
422 * part of the group weight distributed to this span.
424 unsigned long shares;
427 * The sum of all runqueue weights within this span.
429 unsigned long rq_weight;
432 * Weight contributed by tasks; this is the part we can
433 * influence by moving tasks around.
435 unsigned long task_weight;
441 /* Real-Time classes' related field in a runqueue: */
443 struct rt_prio_array active;
444 unsigned long rt_nr_running;
445 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
446 int highest_prio; /* highest queued rt task prio */
449 unsigned long rt_nr_migratory;
455 /* Nests inside the rq lock: */
456 spinlock_t rt_runtime_lock;
458 #ifdef CONFIG_RT_GROUP_SCHED
459 unsigned long rt_nr_boosted;
462 struct list_head leaf_rt_rq_list;
463 struct task_group *tg;
464 struct sched_rt_entity *rt_se;
471 * We add the notion of a root-domain which will be used to define per-domain
472 * variables. Each exclusive cpuset essentially defines an island domain by
473 * fully partitioning the member cpus from any other cpuset. Whenever a new
474 * exclusive cpuset is created, we also create and attach a new root-domain
484 * The "RT overload" flag: it gets set if a CPU has more than
485 * one runnable RT task.
492 * By default the system creates a single root-domain with all cpus as
493 * members (mimicking the global state we have today).
495 static struct root_domain def_root_domain;
500 * This is the main, per-CPU runqueue data structure.
502 * Locking rule: those places that want to lock multiple runqueues
503 * (such as the load balancing or the thread migration code), lock
504 * acquire operations must be ordered by ascending &runqueue.
511 * nr_running and cpu_load should be in the same cacheline because
512 * remote CPUs use both these fields when doing load calculation.
514 unsigned long nr_running;
515 #define CPU_LOAD_IDX_MAX 5
516 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
517 unsigned char idle_at_tick;
519 unsigned long last_tick_seen;
520 unsigned char in_nohz_recently;
522 /* capture load from *all* tasks on this cpu: */
523 struct load_weight load;
524 unsigned long nr_load_updates;
530 #ifdef CONFIG_FAIR_GROUP_SCHED
531 /* list of leaf cfs_rq on this cpu: */
532 struct list_head leaf_cfs_rq_list;
534 #ifdef CONFIG_RT_GROUP_SCHED
535 struct list_head leaf_rt_rq_list;
539 * This is part of a global counter where only the total sum
540 * over all CPUs matters. A task can increase this counter on
541 * one CPU and if it got migrated afterwards it may decrease
542 * it on another CPU. Always updated under the runqueue lock:
544 unsigned long nr_uninterruptible;
546 struct task_struct *curr, *idle;
547 unsigned long next_balance;
548 struct mm_struct *prev_mm;
555 struct root_domain *rd;
556 struct sched_domain *sd;
558 /* For active balancing */
561 /* cpu of this runqueue: */
564 struct task_struct *migration_thread;
565 struct list_head migration_queue;
568 #ifdef CONFIG_SCHED_HRTICK
569 unsigned long hrtick_flags;
570 ktime_t hrtick_expire;
571 struct hrtimer hrtick_timer;
574 #ifdef CONFIG_SCHEDSTATS
576 struct sched_info rq_sched_info;
578 /* sys_sched_yield() stats */
579 unsigned int yld_exp_empty;
580 unsigned int yld_act_empty;
581 unsigned int yld_both_empty;
582 unsigned int yld_count;
584 /* schedule() stats */
585 unsigned int sched_switch;
586 unsigned int sched_count;
587 unsigned int sched_goidle;
589 /* try_to_wake_up() stats */
590 unsigned int ttwu_count;
591 unsigned int ttwu_local;
594 unsigned int bkl_count;
596 struct lock_class_key rq_lock_key;
599 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
601 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
603 rq->curr->sched_class->check_preempt_curr(rq, p);
606 static inline int cpu_of(struct rq *rq)
616 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
617 * See detach_destroy_domains: synchronize_sched for details.
619 * The domain tree of any CPU may only be accessed from within
620 * preempt-disabled sections.
622 #define for_each_domain(cpu, __sd) \
623 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
625 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
626 #define this_rq() (&__get_cpu_var(runqueues))
627 #define task_rq(p) cpu_rq(task_cpu(p))
628 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
630 static inline void update_rq_clock(struct rq *rq)
632 rq->clock = sched_clock_cpu(cpu_of(rq));
636 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
638 #ifdef CONFIG_SCHED_DEBUG
639 # define const_debug __read_mostly
641 # define const_debug static const
645 * Debugging: various feature bits
648 #define SCHED_FEAT(name, enabled) \
649 __SCHED_FEAT_##name ,
652 #include "sched_features.h"
657 #define SCHED_FEAT(name, enabled) \
658 (1UL << __SCHED_FEAT_##name) * enabled |
660 const_debug unsigned int sysctl_sched_features =
661 #include "sched_features.h"
666 #ifdef CONFIG_SCHED_DEBUG
667 #define SCHED_FEAT(name, enabled) \
670 static __read_mostly char *sched_feat_names[] = {
671 #include "sched_features.h"
677 static int sched_feat_open(struct inode *inode, struct file *filp)
679 filp->private_data = inode->i_private;
684 sched_feat_read(struct file *filp, char __user *ubuf,
685 size_t cnt, loff_t *ppos)
692 for (i = 0; sched_feat_names[i]; i++) {
693 len += strlen(sched_feat_names[i]);
697 buf = kmalloc(len + 2, GFP_KERNEL);
701 for (i = 0; sched_feat_names[i]; i++) {
702 if (sysctl_sched_features & (1UL << i))
703 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
705 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
708 r += sprintf(buf + r, "\n");
709 WARN_ON(r >= len + 2);
711 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
719 sched_feat_write(struct file *filp, const char __user *ubuf,
720 size_t cnt, loff_t *ppos)
730 if (copy_from_user(&buf, ubuf, cnt))
735 if (strncmp(buf, "NO_", 3) == 0) {
740 for (i = 0; sched_feat_names[i]; i++) {
741 int len = strlen(sched_feat_names[i]);
743 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
745 sysctl_sched_features &= ~(1UL << i);
747 sysctl_sched_features |= (1UL << i);
752 if (!sched_feat_names[i])
760 static struct file_operations sched_feat_fops = {
761 .open = sched_feat_open,
762 .read = sched_feat_read,
763 .write = sched_feat_write,
766 static __init int sched_init_debug(void)
768 debugfs_create_file("sched_features", 0644, NULL, NULL,
773 late_initcall(sched_init_debug);
777 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
780 * Number of tasks to iterate in a single balance run.
781 * Limited because this is done with IRQs disabled.
783 const_debug unsigned int sysctl_sched_nr_migrate = 32;
786 * period over which we measure -rt task cpu usage in us.
789 unsigned int sysctl_sched_rt_period = 1000000;
791 static __read_mostly int scheduler_running;
794 * part of the period that we allow rt tasks to run in us.
797 int sysctl_sched_rt_runtime = 950000;
799 static inline u64 global_rt_period(void)
801 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
804 static inline u64 global_rt_runtime(void)
806 if (sysctl_sched_rt_period < 0)
809 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
812 unsigned long long time_sync_thresh = 100000;
814 static DEFINE_PER_CPU(unsigned long long, time_offset);
815 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
818 * Global lock which we take every now and then to synchronize
819 * the CPUs time. This method is not warp-safe, but it's good
820 * enough to synchronize slowly diverging time sources and thus
821 * it's good enough for tracing:
823 static DEFINE_SPINLOCK(time_sync_lock);
824 static unsigned long long prev_global_time;
826 static unsigned long long __sync_cpu_clock(unsigned long long time, int cpu)
829 * We want this inlined, to not get tracer function calls
830 * in this critical section:
832 spin_acquire(&time_sync_lock.dep_map, 0, 0, _THIS_IP_);
833 __raw_spin_lock(&time_sync_lock.raw_lock);
835 if (time < prev_global_time) {
836 per_cpu(time_offset, cpu) += prev_global_time - time;
837 time = prev_global_time;
839 prev_global_time = time;
842 __raw_spin_unlock(&time_sync_lock.raw_lock);
843 spin_release(&time_sync_lock.dep_map, 1, _THIS_IP_);
848 static unsigned long long __cpu_clock(int cpu)
850 unsigned long long now;
853 * Only call sched_clock() if the scheduler has already been
854 * initialized (some code might call cpu_clock() very early):
856 if (unlikely(!scheduler_running))
859 now = sched_clock_cpu(cpu);
865 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
866 * clock constructed from sched_clock():
868 unsigned long long cpu_clock(int cpu)
870 unsigned long long prev_cpu_time, time, delta_time;
873 local_irq_save(flags);
874 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
875 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
876 delta_time = time-prev_cpu_time;
878 if (unlikely(delta_time > time_sync_thresh)) {
879 time = __sync_cpu_clock(time, cpu);
880 per_cpu(prev_cpu_time, cpu) = time;
882 local_irq_restore(flags);
886 EXPORT_SYMBOL_GPL(cpu_clock);
888 #ifndef prepare_arch_switch
889 # define prepare_arch_switch(next) do { } while (0)
891 #ifndef finish_arch_switch
892 # define finish_arch_switch(prev) do { } while (0)
895 static inline int task_current(struct rq *rq, struct task_struct *p)
897 return rq->curr == p;
900 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
901 static inline int task_running(struct rq *rq, struct task_struct *p)
903 return task_current(rq, p);
906 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
910 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
912 #ifdef CONFIG_DEBUG_SPINLOCK
913 /* this is a valid case when another task releases the spinlock */
914 rq->lock.owner = current;
917 * If we are tracking spinlock dependencies then we have to
918 * fix up the runqueue lock - which gets 'carried over' from
921 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
923 spin_unlock_irq(&rq->lock);
926 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
927 static inline int task_running(struct rq *rq, struct task_struct *p)
932 return task_current(rq, p);
936 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
940 * We can optimise this out completely for !SMP, because the
941 * SMP rebalancing from interrupt is the only thing that cares
946 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
947 spin_unlock_irq(&rq->lock);
949 spin_unlock(&rq->lock);
953 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
957 * After ->oncpu is cleared, the task can be moved to a different CPU.
958 * We must ensure this doesn't happen until the switch is completely
964 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
968 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
971 * __task_rq_lock - lock the runqueue a given task resides on.
972 * Must be called interrupts disabled.
974 static inline struct rq *__task_rq_lock(struct task_struct *p)
978 struct rq *rq = task_rq(p);
979 spin_lock(&rq->lock);
980 if (likely(rq == task_rq(p)))
982 spin_unlock(&rq->lock);
987 * task_rq_lock - lock the runqueue a given task resides on and disable
988 * interrupts. Note the ordering: we can safely lookup the task_rq without
989 * explicitly disabling preemption.
991 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
997 local_irq_save(*flags);
999 spin_lock(&rq->lock);
1000 if (likely(rq == task_rq(p)))
1002 spin_unlock_irqrestore(&rq->lock, *flags);
1006 static void __task_rq_unlock(struct rq *rq)
1007 __releases(rq->lock)
1009 spin_unlock(&rq->lock);
1012 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1013 __releases(rq->lock)
1015 spin_unlock_irqrestore(&rq->lock, *flags);
1019 * this_rq_lock - lock this runqueue and disable interrupts.
1021 static struct rq *this_rq_lock(void)
1022 __acquires(rq->lock)
1026 local_irq_disable();
1028 spin_lock(&rq->lock);
1033 static void __resched_task(struct task_struct *p, int tif_bit);
1035 static inline void resched_task(struct task_struct *p)
1037 __resched_task(p, TIF_NEED_RESCHED);
1040 #ifdef CONFIG_SCHED_HRTICK
1042 * Use HR-timers to deliver accurate preemption points.
1044 * Its all a bit involved since we cannot program an hrt while holding the
1045 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1048 * When we get rescheduled we reprogram the hrtick_timer outside of the
1051 static inline void resched_hrt(struct task_struct *p)
1053 __resched_task(p, TIF_HRTICK_RESCHED);
1056 static inline void resched_rq(struct rq *rq)
1058 unsigned long flags;
1060 spin_lock_irqsave(&rq->lock, flags);
1061 resched_task(rq->curr);
1062 spin_unlock_irqrestore(&rq->lock, flags);
1066 HRTICK_SET, /* re-programm hrtick_timer */
1067 HRTICK_RESET, /* not a new slice */
1068 HRTICK_BLOCK, /* stop hrtick operations */
1073 * - enabled by features
1074 * - hrtimer is actually high res
1076 static inline int hrtick_enabled(struct rq *rq)
1078 if (!sched_feat(HRTICK))
1080 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
1082 return hrtimer_is_hres_active(&rq->hrtick_timer);
1086 * Called to set the hrtick timer state.
1088 * called with rq->lock held and irqs disabled
1090 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1092 assert_spin_locked(&rq->lock);
1095 * preempt at: now + delay
1098 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1100 * indicate we need to program the timer
1102 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1104 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1107 * New slices are called from the schedule path and don't need a
1108 * forced reschedule.
1111 resched_hrt(rq->curr);
1114 static void hrtick_clear(struct rq *rq)
1116 if (hrtimer_active(&rq->hrtick_timer))
1117 hrtimer_cancel(&rq->hrtick_timer);
1121 * Update the timer from the possible pending state.
1123 static void hrtick_set(struct rq *rq)
1127 unsigned long flags;
1129 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1131 spin_lock_irqsave(&rq->lock, flags);
1132 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1133 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1134 time = rq->hrtick_expire;
1135 clear_thread_flag(TIF_HRTICK_RESCHED);
1136 spin_unlock_irqrestore(&rq->lock, flags);
1139 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1140 if (reset && !hrtimer_active(&rq->hrtick_timer))
1147 * High-resolution timer tick.
1148 * Runs from hardirq context with interrupts disabled.
1150 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1152 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1154 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1156 spin_lock(&rq->lock);
1157 update_rq_clock(rq);
1158 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1159 spin_unlock(&rq->lock);
1161 return HRTIMER_NORESTART;
1164 static void hotplug_hrtick_disable(int cpu)
1166 struct rq *rq = cpu_rq(cpu);
1167 unsigned long flags;
1169 spin_lock_irqsave(&rq->lock, flags);
1170 rq->hrtick_flags = 0;
1171 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1172 spin_unlock_irqrestore(&rq->lock, flags);
1177 static void hotplug_hrtick_enable(int cpu)
1179 struct rq *rq = cpu_rq(cpu);
1180 unsigned long flags;
1182 spin_lock_irqsave(&rq->lock, flags);
1183 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1184 spin_unlock_irqrestore(&rq->lock, flags);
1188 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1190 int cpu = (int)(long)hcpu;
1193 case CPU_UP_CANCELED:
1194 case CPU_UP_CANCELED_FROZEN:
1195 case CPU_DOWN_PREPARE:
1196 case CPU_DOWN_PREPARE_FROZEN:
1198 case CPU_DEAD_FROZEN:
1199 hotplug_hrtick_disable(cpu);
1202 case CPU_UP_PREPARE:
1203 case CPU_UP_PREPARE_FROZEN:
1204 case CPU_DOWN_FAILED:
1205 case CPU_DOWN_FAILED_FROZEN:
1207 case CPU_ONLINE_FROZEN:
1208 hotplug_hrtick_enable(cpu);
1215 static void init_hrtick(void)
1217 hotcpu_notifier(hotplug_hrtick, 0);
1220 static void init_rq_hrtick(struct rq *rq)
1222 rq->hrtick_flags = 0;
1223 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1224 rq->hrtick_timer.function = hrtick;
1225 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1228 void hrtick_resched(void)
1231 unsigned long flags;
1233 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1236 local_irq_save(flags);
1237 rq = cpu_rq(smp_processor_id());
1239 local_irq_restore(flags);
1242 static inline void hrtick_clear(struct rq *rq)
1246 static inline void hrtick_set(struct rq *rq)
1250 static inline void init_rq_hrtick(struct rq *rq)
1254 void hrtick_resched(void)
1258 static inline void init_hrtick(void)
1264 * resched_task - mark a task 'to be rescheduled now'.
1266 * On UP this means the setting of the need_resched flag, on SMP it
1267 * might also involve a cross-CPU call to trigger the scheduler on
1272 #ifndef tsk_is_polling
1273 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1276 static void __resched_task(struct task_struct *p, int tif_bit)
1280 assert_spin_locked(&task_rq(p)->lock);
1282 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1285 set_tsk_thread_flag(p, tif_bit);
1288 if (cpu == smp_processor_id())
1291 /* NEED_RESCHED must be visible before we test polling */
1293 if (!tsk_is_polling(p))
1294 smp_send_reschedule(cpu);
1297 static void resched_cpu(int cpu)
1299 struct rq *rq = cpu_rq(cpu);
1300 unsigned long flags;
1302 if (!spin_trylock_irqsave(&rq->lock, flags))
1304 resched_task(cpu_curr(cpu));
1305 spin_unlock_irqrestore(&rq->lock, flags);
1310 * When add_timer_on() enqueues a timer into the timer wheel of an
1311 * idle CPU then this timer might expire before the next timer event
1312 * which is scheduled to wake up that CPU. In case of a completely
1313 * idle system the next event might even be infinite time into the
1314 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1315 * leaves the inner idle loop so the newly added timer is taken into
1316 * account when the CPU goes back to idle and evaluates the timer
1317 * wheel for the next timer event.
1319 void wake_up_idle_cpu(int cpu)
1321 struct rq *rq = cpu_rq(cpu);
1323 if (cpu == smp_processor_id())
1327 * This is safe, as this function is called with the timer
1328 * wheel base lock of (cpu) held. When the CPU is on the way
1329 * to idle and has not yet set rq->curr to idle then it will
1330 * be serialized on the timer wheel base lock and take the new
1331 * timer into account automatically.
1333 if (rq->curr != rq->idle)
1337 * We can set TIF_RESCHED on the idle task of the other CPU
1338 * lockless. The worst case is that the other CPU runs the
1339 * idle task through an additional NOOP schedule()
1341 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1343 /* NEED_RESCHED must be visible before we test polling */
1345 if (!tsk_is_polling(rq->idle))
1346 smp_send_reschedule(cpu);
1351 static void __resched_task(struct task_struct *p, int tif_bit)
1353 assert_spin_locked(&task_rq(p)->lock);
1354 set_tsk_thread_flag(p, tif_bit);
1358 #if BITS_PER_LONG == 32
1359 # define WMULT_CONST (~0UL)
1361 # define WMULT_CONST (1UL << 32)
1364 #define WMULT_SHIFT 32
1367 * Shift right and round:
1369 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1372 * delta *= weight / lw
1374 static unsigned long
1375 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1376 struct load_weight *lw)
1380 if (!lw->inv_weight)
1381 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)/(lw->weight+1);
1383 tmp = (u64)delta_exec * weight;
1385 * Check whether we'd overflow the 64-bit multiplication:
1387 if (unlikely(tmp > WMULT_CONST))
1388 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1391 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1393 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1396 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1402 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1409 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1410 * of tasks with abnormal "nice" values across CPUs the contribution that
1411 * each task makes to its run queue's load is weighted according to its
1412 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1413 * scaled version of the new time slice allocation that they receive on time
1417 #define WEIGHT_IDLEPRIO 2
1418 #define WMULT_IDLEPRIO (1 << 31)
1421 * Nice levels are multiplicative, with a gentle 10% change for every
1422 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1423 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1424 * that remained on nice 0.
1426 * The "10% effect" is relative and cumulative: from _any_ nice level,
1427 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1428 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1429 * If a task goes up by ~10% and another task goes down by ~10% then
1430 * the relative distance between them is ~25%.)
1432 static const int prio_to_weight[40] = {
1433 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1434 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1435 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1436 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1437 /* 0 */ 1024, 820, 655, 526, 423,
1438 /* 5 */ 335, 272, 215, 172, 137,
1439 /* 10 */ 110, 87, 70, 56, 45,
1440 /* 15 */ 36, 29, 23, 18, 15,
1444 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1446 * In cases where the weight does not change often, we can use the
1447 * precalculated inverse to speed up arithmetics by turning divisions
1448 * into multiplications:
1450 static const u32 prio_to_wmult[40] = {
1451 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1452 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1453 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1454 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1455 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1456 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1457 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1458 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1461 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1464 * runqueue iterator, to support SMP load-balancing between different
1465 * scheduling classes, without having to expose their internal data
1466 * structures to the load-balancing proper:
1468 struct rq_iterator {
1470 struct task_struct *(*start)(void *);
1471 struct task_struct *(*next)(void *);
1475 static unsigned long
1476 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1477 unsigned long max_load_move, struct sched_domain *sd,
1478 enum cpu_idle_type idle, int *all_pinned,
1479 int *this_best_prio, struct rq_iterator *iterator);
1482 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1483 struct sched_domain *sd, enum cpu_idle_type idle,
1484 struct rq_iterator *iterator);
1487 #ifdef CONFIG_CGROUP_CPUACCT
1488 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1490 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1493 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1495 update_load_add(&rq->load, load);
1498 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1500 update_load_sub(&rq->load, load);
1504 static unsigned long source_load(int cpu, int type);
1505 static unsigned long target_load(int cpu, int type);
1506 static unsigned long cpu_avg_load_per_task(int cpu);
1507 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1509 #ifdef CONFIG_FAIR_GROUP_SCHED
1512 * Group load balancing.
1514 * We calculate a few balance domain wide aggregate numbers; load and weight.
1515 * Given the pictures below, and assuming each item has equal weight:
1526 * A and B get 1/3-rd of the total load. C and D get 1/3-rd of A's 1/3-rd,
1527 * which equals 1/9-th of the total load.
1530 * The weight of this group on the selected cpus.
1533 * Direct sum of all the cpu's their rq weight, e.g. A would get 3 while
1537 * Part of the rq_weight contributed by tasks; all groups except B would
1541 static inline struct aggregate_struct *
1542 aggregate(struct task_group *tg, struct sched_domain *sd)
1544 return &tg->cfs_rq[sd->first_cpu]->aggregate;
1547 typedef void (*aggregate_func)(struct task_group *, struct sched_domain *);
1550 * Iterate the full tree, calling @down when first entering a node and @up when
1551 * leaving it for the final time.
1554 void aggregate_walk_tree(aggregate_func down, aggregate_func up,
1555 struct sched_domain *sd)
1557 struct task_group *parent, *child;
1560 parent = &root_task_group;
1562 (*down)(parent, sd);
1563 list_for_each_entry_rcu(child, &parent->children, siblings) {
1573 parent = parent->parent;
1580 * Calculate the aggregate runqueue weight.
1583 void aggregate_group_weight(struct task_group *tg, struct sched_domain *sd)
1585 unsigned long rq_weight = 0;
1586 unsigned long task_weight = 0;
1589 for_each_cpu_mask(i, sd->span) {
1590 rq_weight += tg->cfs_rq[i]->load.weight;
1591 task_weight += tg->cfs_rq[i]->task_weight;
1594 aggregate(tg, sd)->rq_weight = rq_weight;
1595 aggregate(tg, sd)->task_weight = task_weight;
1599 * Compute the weight of this group on the given cpus.
1602 void aggregate_group_shares(struct task_group *tg, struct sched_domain *sd)
1604 unsigned long shares = 0;
1607 for_each_cpu_mask(i, sd->span)
1608 shares += tg->cfs_rq[i]->shares;
1610 if ((!shares && aggregate(tg, sd)->rq_weight) || shares > tg->shares)
1611 shares = tg->shares;
1613 aggregate(tg, sd)->shares = shares;
1617 * Compute the load fraction assigned to this group, relies on the aggregate
1618 * weight and this group's parent's load, i.e. top-down.
1621 void aggregate_group_load(struct task_group *tg, struct sched_domain *sd)
1629 for_each_cpu_mask(i, sd->span)
1630 load += cpu_rq(i)->load.weight;
1633 load = aggregate(tg->parent, sd)->load;
1636 * shares is our weight in the parent's rq so
1637 * shares/parent->rq_weight gives our fraction of the load
1639 load *= aggregate(tg, sd)->shares;
1640 load /= aggregate(tg->parent, sd)->rq_weight + 1;
1643 aggregate(tg, sd)->load = load;
1646 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1649 * Calculate and set the cpu's group shares.
1652 __update_group_shares_cpu(struct task_group *tg, struct sched_domain *sd,
1656 unsigned long shares;
1657 unsigned long rq_weight;
1662 rq_weight = tg->cfs_rq[tcpu]->load.weight;
1665 * If there are currently no tasks on the cpu pretend there is one of
1666 * average load so that when a new task gets to run here it will not
1667 * get delayed by group starvation.
1671 rq_weight = NICE_0_LOAD;
1675 * \Sum shares * rq_weight
1676 * shares = -----------------------
1680 shares = aggregate(tg, sd)->shares * rq_weight;
1681 shares /= aggregate(tg, sd)->rq_weight + 1;
1684 * record the actual number of shares, not the boosted amount.
1686 tg->cfs_rq[tcpu]->shares = boost ? 0 : shares;
1688 if (shares < MIN_SHARES)
1689 shares = MIN_SHARES;
1690 else if (shares > MAX_SHARES)
1691 shares = MAX_SHARES;
1693 __set_se_shares(tg->se[tcpu], shares);
1697 * Re-adjust the weights on the cpu the task came from and on the cpu the
1701 __move_group_shares(struct task_group *tg, struct sched_domain *sd,
1704 unsigned long shares;
1706 shares = tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1708 __update_group_shares_cpu(tg, sd, scpu);
1709 __update_group_shares_cpu(tg, sd, dcpu);
1712 * ensure we never loose shares due to rounding errors in the
1713 * above redistribution.
1715 shares -= tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1717 tg->cfs_rq[dcpu]->shares += shares;
1721 * Because changing a group's shares changes the weight of the super-group
1722 * we need to walk up the tree and change all shares until we hit the root.
1725 move_group_shares(struct task_group *tg, struct sched_domain *sd,
1729 __move_group_shares(tg, sd, scpu, dcpu);
1735 void aggregate_group_set_shares(struct task_group *tg, struct sched_domain *sd)
1737 unsigned long shares = aggregate(tg, sd)->shares;
1740 for_each_cpu_mask(i, sd->span) {
1741 struct rq *rq = cpu_rq(i);
1742 unsigned long flags;
1744 spin_lock_irqsave(&rq->lock, flags);
1745 __update_group_shares_cpu(tg, sd, i);
1746 spin_unlock_irqrestore(&rq->lock, flags);
1749 aggregate_group_shares(tg, sd);
1752 * ensure we never loose shares due to rounding errors in the
1753 * above redistribution.
1755 shares -= aggregate(tg, sd)->shares;
1757 tg->cfs_rq[sd->first_cpu]->shares += shares;
1758 aggregate(tg, sd)->shares += shares;
1763 * Calculate the accumulative weight and recursive load of each task group
1764 * while walking down the tree.
1767 void aggregate_get_down(struct task_group *tg, struct sched_domain *sd)
1769 aggregate_group_weight(tg, sd);
1770 aggregate_group_shares(tg, sd);
1771 aggregate_group_load(tg, sd);
1775 * Rebalance the cpu shares while walking back up the tree.
1778 void aggregate_get_up(struct task_group *tg, struct sched_domain *sd)
1780 aggregate_group_set_shares(tg, sd);
1783 static DEFINE_PER_CPU(spinlock_t, aggregate_lock);
1785 static void __init init_aggregate(void)
1789 for_each_possible_cpu(i)
1790 spin_lock_init(&per_cpu(aggregate_lock, i));
1793 static int get_aggregate(struct sched_domain *sd)
1795 if (!spin_trylock(&per_cpu(aggregate_lock, sd->first_cpu)))
1798 aggregate_walk_tree(aggregate_get_down, aggregate_get_up, sd);
1802 static void put_aggregate(struct sched_domain *sd)
1804 spin_unlock(&per_cpu(aggregate_lock, sd->first_cpu));
1807 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1809 cfs_rq->shares = shares;
1814 static inline void init_aggregate(void)
1818 static inline int get_aggregate(struct sched_domain *sd)
1823 static inline void put_aggregate(struct sched_domain *sd)
1828 #else /* CONFIG_SMP */
1830 #ifdef CONFIG_FAIR_GROUP_SCHED
1831 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1836 #endif /* CONFIG_SMP */
1838 #include "sched_stats.h"
1839 #include "sched_idletask.c"
1840 #include "sched_fair.c"
1841 #include "sched_rt.c"
1842 #ifdef CONFIG_SCHED_DEBUG
1843 # include "sched_debug.c"
1846 #define sched_class_highest (&rt_sched_class)
1848 static void inc_nr_running(struct rq *rq)
1853 static void dec_nr_running(struct rq *rq)
1858 static void set_load_weight(struct task_struct *p)
1860 if (task_has_rt_policy(p)) {
1861 p->se.load.weight = prio_to_weight[0] * 2;
1862 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1867 * SCHED_IDLE tasks get minimal weight:
1869 if (p->policy == SCHED_IDLE) {
1870 p->se.load.weight = WEIGHT_IDLEPRIO;
1871 p->se.load.inv_weight = WMULT_IDLEPRIO;
1875 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1876 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1879 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1881 sched_info_queued(p);
1882 p->sched_class->enqueue_task(rq, p, wakeup);
1886 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1888 p->sched_class->dequeue_task(rq, p, sleep);
1893 * __normal_prio - return the priority that is based on the static prio
1895 static inline int __normal_prio(struct task_struct *p)
1897 return p->static_prio;
1901 * Calculate the expected normal priority: i.e. priority
1902 * without taking RT-inheritance into account. Might be
1903 * boosted by interactivity modifiers. Changes upon fork,
1904 * setprio syscalls, and whenever the interactivity
1905 * estimator recalculates.
1907 static inline int normal_prio(struct task_struct *p)
1911 if (task_has_rt_policy(p))
1912 prio = MAX_RT_PRIO-1 - p->rt_priority;
1914 prio = __normal_prio(p);
1919 * Calculate the current priority, i.e. the priority
1920 * taken into account by the scheduler. This value might
1921 * be boosted by RT tasks, or might be boosted by
1922 * interactivity modifiers. Will be RT if the task got
1923 * RT-boosted. If not then it returns p->normal_prio.
1925 static int effective_prio(struct task_struct *p)
1927 p->normal_prio = normal_prio(p);
1929 * If we are RT tasks or we were boosted to RT priority,
1930 * keep the priority unchanged. Otherwise, update priority
1931 * to the normal priority:
1933 if (!rt_prio(p->prio))
1934 return p->normal_prio;
1939 * activate_task - move a task to the runqueue.
1941 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1943 if (task_contributes_to_load(p))
1944 rq->nr_uninterruptible--;
1946 enqueue_task(rq, p, wakeup);
1951 * deactivate_task - remove a task from the runqueue.
1953 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1955 if (task_contributes_to_load(p))
1956 rq->nr_uninterruptible++;
1958 dequeue_task(rq, p, sleep);
1963 * task_curr - is this task currently executing on a CPU?
1964 * @p: the task in question.
1966 inline int task_curr(const struct task_struct *p)
1968 return cpu_curr(task_cpu(p)) == p;
1971 /* Used instead of source_load when we know the type == 0 */
1972 unsigned long weighted_cpuload(const int cpu)
1974 return cpu_rq(cpu)->load.weight;
1977 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1979 set_task_rq(p, cpu);
1982 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1983 * successfuly executed on another CPU. We must ensure that updates of
1984 * per-task data have been completed by this moment.
1987 task_thread_info(p)->cpu = cpu;
1991 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1992 const struct sched_class *prev_class,
1993 int oldprio, int running)
1995 if (prev_class != p->sched_class) {
1996 if (prev_class->switched_from)
1997 prev_class->switched_from(rq, p, running);
1998 p->sched_class->switched_to(rq, p, running);
2000 p->sched_class->prio_changed(rq, p, oldprio, running);
2006 * Is this task likely cache-hot:
2009 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2014 * Buddy candidates are cache hot:
2016 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
2019 if (p->sched_class != &fair_sched_class)
2022 if (sysctl_sched_migration_cost == -1)
2024 if (sysctl_sched_migration_cost == 0)
2027 delta = now - p->se.exec_start;
2029 return delta < (s64)sysctl_sched_migration_cost;
2033 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2035 int old_cpu = task_cpu(p);
2036 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2037 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2038 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2041 clock_offset = old_rq->clock - new_rq->clock;
2043 #ifdef CONFIG_SCHEDSTATS
2044 if (p->se.wait_start)
2045 p->se.wait_start -= clock_offset;
2046 if (p->se.sleep_start)
2047 p->se.sleep_start -= clock_offset;
2048 if (p->se.block_start)
2049 p->se.block_start -= clock_offset;
2050 if (old_cpu != new_cpu) {
2051 schedstat_inc(p, se.nr_migrations);
2052 if (task_hot(p, old_rq->clock, NULL))
2053 schedstat_inc(p, se.nr_forced2_migrations);
2056 p->se.vruntime -= old_cfsrq->min_vruntime -
2057 new_cfsrq->min_vruntime;
2059 __set_task_cpu(p, new_cpu);
2062 struct migration_req {
2063 struct list_head list;
2065 struct task_struct *task;
2068 struct completion done;
2072 * The task's runqueue lock must be held.
2073 * Returns true if you have to wait for migration thread.
2076 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2078 struct rq *rq = task_rq(p);
2081 * If the task is not on a runqueue (and not running), then
2082 * it is sufficient to simply update the task's cpu field.
2084 if (!p->se.on_rq && !task_running(rq, p)) {
2085 set_task_cpu(p, dest_cpu);
2089 init_completion(&req->done);
2091 req->dest_cpu = dest_cpu;
2092 list_add(&req->list, &rq->migration_queue);
2098 * wait_task_inactive - wait for a thread to unschedule.
2100 * The caller must ensure that the task *will* unschedule sometime soon,
2101 * else this function might spin for a *long* time. This function can't
2102 * be called with interrupts off, or it may introduce deadlock with
2103 * smp_call_function() if an IPI is sent by the same process we are
2104 * waiting to become inactive.
2106 void wait_task_inactive(struct task_struct *p)
2108 unsigned long flags;
2114 * We do the initial early heuristics without holding
2115 * any task-queue locks at all. We'll only try to get
2116 * the runqueue lock when things look like they will
2122 * If the task is actively running on another CPU
2123 * still, just relax and busy-wait without holding
2126 * NOTE! Since we don't hold any locks, it's not
2127 * even sure that "rq" stays as the right runqueue!
2128 * But we don't care, since "task_running()" will
2129 * return false if the runqueue has changed and p
2130 * is actually now running somewhere else!
2132 while (task_running(rq, p))
2136 * Ok, time to look more closely! We need the rq
2137 * lock now, to be *sure*. If we're wrong, we'll
2138 * just go back and repeat.
2140 rq = task_rq_lock(p, &flags);
2141 running = task_running(rq, p);
2142 on_rq = p->se.on_rq;
2143 task_rq_unlock(rq, &flags);
2146 * Was it really running after all now that we
2147 * checked with the proper locks actually held?
2149 * Oops. Go back and try again..
2151 if (unlikely(running)) {
2157 * It's not enough that it's not actively running,
2158 * it must be off the runqueue _entirely_, and not
2161 * So if it wa still runnable (but just not actively
2162 * running right now), it's preempted, and we should
2163 * yield - it could be a while.
2165 if (unlikely(on_rq)) {
2166 schedule_timeout_uninterruptible(1);
2171 * Ahh, all good. It wasn't running, and it wasn't
2172 * runnable, which means that it will never become
2173 * running in the future either. We're all done!
2180 * kick_process - kick a running thread to enter/exit the kernel
2181 * @p: the to-be-kicked thread
2183 * Cause a process which is running on another CPU to enter
2184 * kernel-mode, without any delay. (to get signals handled.)
2186 * NOTE: this function doesnt have to take the runqueue lock,
2187 * because all it wants to ensure is that the remote task enters
2188 * the kernel. If the IPI races and the task has been migrated
2189 * to another CPU then no harm is done and the purpose has been
2192 void kick_process(struct task_struct *p)
2198 if ((cpu != smp_processor_id()) && task_curr(p))
2199 smp_send_reschedule(cpu);
2204 * Return a low guess at the load of a migration-source cpu weighted
2205 * according to the scheduling class and "nice" value.
2207 * We want to under-estimate the load of migration sources, to
2208 * balance conservatively.
2210 static unsigned long source_load(int cpu, int type)
2212 struct rq *rq = cpu_rq(cpu);
2213 unsigned long total = weighted_cpuload(cpu);
2218 return min(rq->cpu_load[type-1], total);
2222 * Return a high guess at the load of a migration-target cpu weighted
2223 * according to the scheduling class and "nice" value.
2225 static unsigned long target_load(int cpu, int type)
2227 struct rq *rq = cpu_rq(cpu);
2228 unsigned long total = weighted_cpuload(cpu);
2233 return max(rq->cpu_load[type-1], total);
2237 * Return the average load per task on the cpu's run queue
2239 static unsigned long cpu_avg_load_per_task(int cpu)
2241 struct rq *rq = cpu_rq(cpu);
2242 unsigned long total = weighted_cpuload(cpu);
2243 unsigned long n = rq->nr_running;
2245 return n ? total / n : SCHED_LOAD_SCALE;
2249 * find_idlest_group finds and returns the least busy CPU group within the
2252 static struct sched_group *
2253 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2255 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2256 unsigned long min_load = ULONG_MAX, this_load = 0;
2257 int load_idx = sd->forkexec_idx;
2258 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2261 unsigned long load, avg_load;
2265 /* Skip over this group if it has no CPUs allowed */
2266 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2269 local_group = cpu_isset(this_cpu, group->cpumask);
2271 /* Tally up the load of all CPUs in the group */
2274 for_each_cpu_mask(i, group->cpumask) {
2275 /* Bias balancing toward cpus of our domain */
2277 load = source_load(i, load_idx);
2279 load = target_load(i, load_idx);
2284 /* Adjust by relative CPU power of the group */
2285 avg_load = sg_div_cpu_power(group,
2286 avg_load * SCHED_LOAD_SCALE);
2289 this_load = avg_load;
2291 } else if (avg_load < min_load) {
2292 min_load = avg_load;
2295 } while (group = group->next, group != sd->groups);
2297 if (!idlest || 100*this_load < imbalance*min_load)
2303 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2306 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2309 unsigned long load, min_load = ULONG_MAX;
2313 /* Traverse only the allowed CPUs */
2314 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2316 for_each_cpu_mask(i, *tmp) {
2317 load = weighted_cpuload(i);
2319 if (load < min_load || (load == min_load && i == this_cpu)) {
2329 * sched_balance_self: balance the current task (running on cpu) in domains
2330 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2333 * Balance, ie. select the least loaded group.
2335 * Returns the target CPU number, or the same CPU if no balancing is needed.
2337 * preempt must be disabled.
2339 static int sched_balance_self(int cpu, int flag)
2341 struct task_struct *t = current;
2342 struct sched_domain *tmp, *sd = NULL;
2344 for_each_domain(cpu, tmp) {
2346 * If power savings logic is enabled for a domain, stop there.
2348 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2350 if (tmp->flags & flag)
2355 cpumask_t span, tmpmask;
2356 struct sched_group *group;
2357 int new_cpu, weight;
2359 if (!(sd->flags & flag)) {
2365 group = find_idlest_group(sd, t, cpu);
2371 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2372 if (new_cpu == -1 || new_cpu == cpu) {
2373 /* Now try balancing at a lower domain level of cpu */
2378 /* Now try balancing at a lower domain level of new_cpu */
2381 weight = cpus_weight(span);
2382 for_each_domain(cpu, tmp) {
2383 if (weight <= cpus_weight(tmp->span))
2385 if (tmp->flags & flag)
2388 /* while loop will break here if sd == NULL */
2394 #endif /* CONFIG_SMP */
2397 * try_to_wake_up - wake up a thread
2398 * @p: the to-be-woken-up thread
2399 * @state: the mask of task states that can be woken
2400 * @sync: do a synchronous wakeup?
2402 * Put it on the run-queue if it's not already there. The "current"
2403 * thread is always on the run-queue (except when the actual
2404 * re-schedule is in progress), and as such you're allowed to do
2405 * the simpler "current->state = TASK_RUNNING" to mark yourself
2406 * runnable without the overhead of this.
2408 * returns failure only if the task is already active.
2410 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2412 int cpu, orig_cpu, this_cpu, success = 0;
2413 unsigned long flags;
2417 if (!sched_feat(SYNC_WAKEUPS))
2421 rq = task_rq_lock(p, &flags);
2422 old_state = p->state;
2423 if (!(old_state & state))
2431 this_cpu = smp_processor_id();
2434 if (unlikely(task_running(rq, p)))
2437 cpu = p->sched_class->select_task_rq(p, sync);
2438 if (cpu != orig_cpu) {
2439 set_task_cpu(p, cpu);
2440 task_rq_unlock(rq, &flags);
2441 /* might preempt at this point */
2442 rq = task_rq_lock(p, &flags);
2443 old_state = p->state;
2444 if (!(old_state & state))
2449 this_cpu = smp_processor_id();
2453 #ifdef CONFIG_SCHEDSTATS
2454 schedstat_inc(rq, ttwu_count);
2455 if (cpu == this_cpu)
2456 schedstat_inc(rq, ttwu_local);
2458 struct sched_domain *sd;
2459 for_each_domain(this_cpu, sd) {
2460 if (cpu_isset(cpu, sd->span)) {
2461 schedstat_inc(sd, ttwu_wake_remote);
2469 #endif /* CONFIG_SMP */
2470 ftrace_wake_up_task(p, rq->curr);
2471 schedstat_inc(p, se.nr_wakeups);
2473 schedstat_inc(p, se.nr_wakeups_sync);
2474 if (orig_cpu != cpu)
2475 schedstat_inc(p, se.nr_wakeups_migrate);
2476 if (cpu == this_cpu)
2477 schedstat_inc(p, se.nr_wakeups_local);
2479 schedstat_inc(p, se.nr_wakeups_remote);
2480 update_rq_clock(rq);
2481 activate_task(rq, p, 1);
2485 check_preempt_curr(rq, p);
2487 p->state = TASK_RUNNING;
2489 if (p->sched_class->task_wake_up)
2490 p->sched_class->task_wake_up(rq, p);
2493 task_rq_unlock(rq, &flags);
2498 int wake_up_process(struct task_struct *p)
2500 return try_to_wake_up(p, TASK_ALL, 0);
2502 EXPORT_SYMBOL(wake_up_process);
2504 int wake_up_state(struct task_struct *p, unsigned int state)
2506 return try_to_wake_up(p, state, 0);
2510 * Perform scheduler related setup for a newly forked process p.
2511 * p is forked by current.
2513 * __sched_fork() is basic setup used by init_idle() too:
2515 static void __sched_fork(struct task_struct *p)
2517 p->se.exec_start = 0;
2518 p->se.sum_exec_runtime = 0;
2519 p->se.prev_sum_exec_runtime = 0;
2520 p->se.last_wakeup = 0;
2521 p->se.avg_overlap = 0;
2523 #ifdef CONFIG_SCHEDSTATS
2524 p->se.wait_start = 0;
2525 p->se.sum_sleep_runtime = 0;
2526 p->se.sleep_start = 0;
2527 p->se.block_start = 0;
2528 p->se.sleep_max = 0;
2529 p->se.block_max = 0;
2531 p->se.slice_max = 0;
2535 INIT_LIST_HEAD(&p->rt.run_list);
2537 INIT_LIST_HEAD(&p->se.group_node);
2539 #ifdef CONFIG_PREEMPT_NOTIFIERS
2540 INIT_HLIST_HEAD(&p->preempt_notifiers);
2544 * We mark the process as running here, but have not actually
2545 * inserted it onto the runqueue yet. This guarantees that
2546 * nobody will actually run it, and a signal or other external
2547 * event cannot wake it up and insert it on the runqueue either.
2549 p->state = TASK_RUNNING;
2553 * fork()/clone()-time setup:
2555 void sched_fork(struct task_struct *p, int clone_flags)
2557 int cpu = get_cpu();
2562 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2564 set_task_cpu(p, cpu);
2567 * Make sure we do not leak PI boosting priority to the child:
2569 p->prio = current->normal_prio;
2570 if (!rt_prio(p->prio))
2571 p->sched_class = &fair_sched_class;
2573 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2574 if (likely(sched_info_on()))
2575 memset(&p->sched_info, 0, sizeof(p->sched_info));
2577 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2580 #ifdef CONFIG_PREEMPT
2581 /* Want to start with kernel preemption disabled. */
2582 task_thread_info(p)->preempt_count = 1;
2588 * wake_up_new_task - wake up a newly created task for the first time.
2590 * This function will do some initial scheduler statistics housekeeping
2591 * that must be done for every newly created context, then puts the task
2592 * on the runqueue and wakes it.
2594 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2596 unsigned long flags;
2599 rq = task_rq_lock(p, &flags);
2600 BUG_ON(p->state != TASK_RUNNING);
2601 update_rq_clock(rq);
2603 p->prio = effective_prio(p);
2605 if (!p->sched_class->task_new || !current->se.on_rq) {
2606 activate_task(rq, p, 0);
2609 * Let the scheduling class do new task startup
2610 * management (if any):
2612 p->sched_class->task_new(rq, p);
2615 ftrace_wake_up_new_task(p, rq->curr);
2616 check_preempt_curr(rq, p);
2618 if (p->sched_class->task_wake_up)
2619 p->sched_class->task_wake_up(rq, p);
2621 task_rq_unlock(rq, &flags);
2624 #ifdef CONFIG_PREEMPT_NOTIFIERS
2627 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2628 * @notifier: notifier struct to register
2630 void preempt_notifier_register(struct preempt_notifier *notifier)
2632 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2634 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2637 * preempt_notifier_unregister - no longer interested in preemption notifications
2638 * @notifier: notifier struct to unregister
2640 * This is safe to call from within a preemption notifier.
2642 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2644 hlist_del(¬ifier->link);
2646 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2648 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2650 struct preempt_notifier *notifier;
2651 struct hlist_node *node;
2653 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2654 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2658 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2659 struct task_struct *next)
2661 struct preempt_notifier *notifier;
2662 struct hlist_node *node;
2664 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2665 notifier->ops->sched_out(notifier, next);
2670 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2675 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2676 struct task_struct *next)
2683 * prepare_task_switch - prepare to switch tasks
2684 * @rq: the runqueue preparing to switch
2685 * @prev: the current task that is being switched out
2686 * @next: the task we are going to switch to.
2688 * This is called with the rq lock held and interrupts off. It must
2689 * be paired with a subsequent finish_task_switch after the context
2692 * prepare_task_switch sets up locking and calls architecture specific
2696 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2697 struct task_struct *next)
2699 fire_sched_out_preempt_notifiers(prev, next);
2700 prepare_lock_switch(rq, next);
2701 prepare_arch_switch(next);
2705 * finish_task_switch - clean up after a task-switch
2706 * @rq: runqueue associated with task-switch
2707 * @prev: the thread we just switched away from.
2709 * finish_task_switch must be called after the context switch, paired
2710 * with a prepare_task_switch call before the context switch.
2711 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2712 * and do any other architecture-specific cleanup actions.
2714 * Note that we may have delayed dropping an mm in context_switch(). If
2715 * so, we finish that here outside of the runqueue lock. (Doing it
2716 * with the lock held can cause deadlocks; see schedule() for
2719 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2720 __releases(rq->lock)
2722 struct mm_struct *mm = rq->prev_mm;
2728 * A task struct has one reference for the use as "current".
2729 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2730 * schedule one last time. The schedule call will never return, and
2731 * the scheduled task must drop that reference.
2732 * The test for TASK_DEAD must occur while the runqueue locks are
2733 * still held, otherwise prev could be scheduled on another cpu, die
2734 * there before we look at prev->state, and then the reference would
2736 * Manfred Spraul <manfred@colorfullife.com>
2738 prev_state = prev->state;
2739 finish_arch_switch(prev);
2740 finish_lock_switch(rq, prev);
2742 if (current->sched_class->post_schedule)
2743 current->sched_class->post_schedule(rq);
2746 fire_sched_in_preempt_notifiers(current);
2749 if (unlikely(prev_state == TASK_DEAD)) {
2751 * Remove function-return probe instances associated with this
2752 * task and put them back on the free list.
2754 kprobe_flush_task(prev);
2755 put_task_struct(prev);
2760 * schedule_tail - first thing a freshly forked thread must call.
2761 * @prev: the thread we just switched away from.
2763 asmlinkage void schedule_tail(struct task_struct *prev)
2764 __releases(rq->lock)
2766 struct rq *rq = this_rq();
2768 finish_task_switch(rq, prev);
2769 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2770 /* In this case, finish_task_switch does not reenable preemption */
2773 if (current->set_child_tid)
2774 put_user(task_pid_vnr(current), current->set_child_tid);
2778 * context_switch - switch to the new MM and the new
2779 * thread's register state.
2782 context_switch(struct rq *rq, struct task_struct *prev,
2783 struct task_struct *next)
2785 struct mm_struct *mm, *oldmm;
2787 prepare_task_switch(rq, prev, next);
2788 ftrace_ctx_switch(prev, next);
2790 oldmm = prev->active_mm;
2792 * For paravirt, this is coupled with an exit in switch_to to
2793 * combine the page table reload and the switch backend into
2796 arch_enter_lazy_cpu_mode();
2798 if (unlikely(!mm)) {
2799 next->active_mm = oldmm;
2800 atomic_inc(&oldmm->mm_count);
2801 enter_lazy_tlb(oldmm, next);
2803 switch_mm(oldmm, mm, next);
2805 if (unlikely(!prev->mm)) {
2806 prev->active_mm = NULL;
2807 rq->prev_mm = oldmm;
2810 * Since the runqueue lock will be released by the next
2811 * task (which is an invalid locking op but in the case
2812 * of the scheduler it's an obvious special-case), so we
2813 * do an early lockdep release here:
2815 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2816 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2819 /* Here we just switch the register state and the stack. */
2820 switch_to(prev, next, prev);
2824 * this_rq must be evaluated again because prev may have moved
2825 * CPUs since it called schedule(), thus the 'rq' on its stack
2826 * frame will be invalid.
2828 finish_task_switch(this_rq(), prev);
2832 * nr_running, nr_uninterruptible and nr_context_switches:
2834 * externally visible scheduler statistics: current number of runnable
2835 * threads, current number of uninterruptible-sleeping threads, total
2836 * number of context switches performed since bootup.
2838 unsigned long nr_running(void)
2840 unsigned long i, sum = 0;
2842 for_each_online_cpu(i)
2843 sum += cpu_rq(i)->nr_running;
2848 unsigned long nr_uninterruptible(void)
2850 unsigned long i, sum = 0;
2852 for_each_possible_cpu(i)
2853 sum += cpu_rq(i)->nr_uninterruptible;
2856 * Since we read the counters lockless, it might be slightly
2857 * inaccurate. Do not allow it to go below zero though:
2859 if (unlikely((long)sum < 0))
2865 unsigned long long nr_context_switches(void)
2868 unsigned long long sum = 0;
2870 for_each_possible_cpu(i)
2871 sum += cpu_rq(i)->nr_switches;
2876 unsigned long nr_iowait(void)
2878 unsigned long i, sum = 0;
2880 for_each_possible_cpu(i)
2881 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2886 unsigned long nr_active(void)
2888 unsigned long i, running = 0, uninterruptible = 0;
2890 for_each_online_cpu(i) {
2891 running += cpu_rq(i)->nr_running;
2892 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2895 if (unlikely((long)uninterruptible < 0))
2896 uninterruptible = 0;
2898 return running + uninterruptible;
2902 * Update rq->cpu_load[] statistics. This function is usually called every
2903 * scheduler tick (TICK_NSEC).
2905 static void update_cpu_load(struct rq *this_rq)
2907 unsigned long this_load = this_rq->load.weight;
2910 this_rq->nr_load_updates++;
2912 /* Update our load: */
2913 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2914 unsigned long old_load, new_load;
2916 /* scale is effectively 1 << i now, and >> i divides by scale */
2918 old_load = this_rq->cpu_load[i];
2919 new_load = this_load;
2921 * Round up the averaging division if load is increasing. This
2922 * prevents us from getting stuck on 9 if the load is 10, for
2925 if (new_load > old_load)
2926 new_load += scale-1;
2927 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2934 * double_rq_lock - safely lock two runqueues
2936 * Note this does not disable interrupts like task_rq_lock,
2937 * you need to do so manually before calling.
2939 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2940 __acquires(rq1->lock)
2941 __acquires(rq2->lock)
2943 BUG_ON(!irqs_disabled());
2945 spin_lock(&rq1->lock);
2946 __acquire(rq2->lock); /* Fake it out ;) */
2949 spin_lock(&rq1->lock);
2950 spin_lock(&rq2->lock);
2952 spin_lock(&rq2->lock);
2953 spin_lock(&rq1->lock);
2956 update_rq_clock(rq1);
2957 update_rq_clock(rq2);
2961 * double_rq_unlock - safely unlock two runqueues
2963 * Note this does not restore interrupts like task_rq_unlock,
2964 * you need to do so manually after calling.
2966 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2967 __releases(rq1->lock)
2968 __releases(rq2->lock)
2970 spin_unlock(&rq1->lock);
2972 spin_unlock(&rq2->lock);
2974 __release(rq2->lock);
2978 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2980 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2981 __releases(this_rq->lock)
2982 __acquires(busiest->lock)
2983 __acquires(this_rq->lock)
2987 if (unlikely(!irqs_disabled())) {
2988 /* printk() doesn't work good under rq->lock */
2989 spin_unlock(&this_rq->lock);
2992 if (unlikely(!spin_trylock(&busiest->lock))) {
2993 if (busiest < this_rq) {
2994 spin_unlock(&this_rq->lock);
2995 spin_lock(&busiest->lock);
2996 spin_lock(&this_rq->lock);
2999 spin_lock(&busiest->lock);
3005 * If dest_cpu is allowed for this process, migrate the task to it.
3006 * This is accomplished by forcing the cpu_allowed mask to only
3007 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3008 * the cpu_allowed mask is restored.
3010 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3012 struct migration_req req;
3013 unsigned long flags;
3016 rq = task_rq_lock(p, &flags);
3017 if (!cpu_isset(dest_cpu, p->cpus_allowed)
3018 || unlikely(cpu_is_offline(dest_cpu)))
3021 /* force the process onto the specified CPU */
3022 if (migrate_task(p, dest_cpu, &req)) {
3023 /* Need to wait for migration thread (might exit: take ref). */
3024 struct task_struct *mt = rq->migration_thread;
3026 get_task_struct(mt);
3027 task_rq_unlock(rq, &flags);
3028 wake_up_process(mt);
3029 put_task_struct(mt);
3030 wait_for_completion(&req.done);
3035 task_rq_unlock(rq, &flags);
3039 * sched_exec - execve() is a valuable balancing opportunity, because at
3040 * this point the task has the smallest effective memory and cache footprint.
3042 void sched_exec(void)
3044 int new_cpu, this_cpu = get_cpu();
3045 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3047 if (new_cpu != this_cpu)
3048 sched_migrate_task(current, new_cpu);
3052 * pull_task - move a task from a remote runqueue to the local runqueue.
3053 * Both runqueues must be locked.
3055 static void pull_task(struct rq *src_rq, struct task_struct *p,
3056 struct rq *this_rq, int this_cpu)
3058 deactivate_task(src_rq, p, 0);
3059 set_task_cpu(p, this_cpu);
3060 activate_task(this_rq, p, 0);
3062 * Note that idle threads have a prio of MAX_PRIO, for this test
3063 * to be always true for them.
3065 check_preempt_curr(this_rq, p);
3069 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3072 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3073 struct sched_domain *sd, enum cpu_idle_type idle,
3077 * We do not migrate tasks that are:
3078 * 1) running (obviously), or
3079 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3080 * 3) are cache-hot on their current CPU.
3082 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
3083 schedstat_inc(p, se.nr_failed_migrations_affine);
3088 if (task_running(rq, p)) {
3089 schedstat_inc(p, se.nr_failed_migrations_running);
3094 * Aggressive migration if:
3095 * 1) task is cache cold, or
3096 * 2) too many balance attempts have failed.
3099 if (!task_hot(p, rq->clock, sd) ||
3100 sd->nr_balance_failed > sd->cache_nice_tries) {
3101 #ifdef CONFIG_SCHEDSTATS
3102 if (task_hot(p, rq->clock, sd)) {
3103 schedstat_inc(sd, lb_hot_gained[idle]);
3104 schedstat_inc(p, se.nr_forced_migrations);
3110 if (task_hot(p, rq->clock, sd)) {
3111 schedstat_inc(p, se.nr_failed_migrations_hot);
3117 static unsigned long
3118 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3119 unsigned long max_load_move, struct sched_domain *sd,
3120 enum cpu_idle_type idle, int *all_pinned,
3121 int *this_best_prio, struct rq_iterator *iterator)
3123 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
3124 struct task_struct *p;
3125 long rem_load_move = max_load_move;
3127 if (max_load_move == 0)
3133 * Start the load-balancing iterator:
3135 p = iterator->start(iterator->arg);
3137 if (!p || loops++ > sysctl_sched_nr_migrate)
3140 * To help distribute high priority tasks across CPUs we don't
3141 * skip a task if it will be the highest priority task (i.e. smallest
3142 * prio value) on its new queue regardless of its load weight
3144 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
3145 SCHED_LOAD_SCALE_FUZZ;
3146 if ((skip_for_load && p->prio >= *this_best_prio) ||
3147 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3148 p = iterator->next(iterator->arg);
3152 pull_task(busiest, p, this_rq, this_cpu);
3154 rem_load_move -= p->se.load.weight;
3157 * We only want to steal up to the prescribed amount of weighted load.
3159 if (rem_load_move > 0) {
3160 if (p->prio < *this_best_prio)
3161 *this_best_prio = p->prio;
3162 p = iterator->next(iterator->arg);
3167 * Right now, this is one of only two places pull_task() is called,
3168 * so we can safely collect pull_task() stats here rather than
3169 * inside pull_task().
3171 schedstat_add(sd, lb_gained[idle], pulled);
3174 *all_pinned = pinned;
3176 return max_load_move - rem_load_move;
3180 * move_tasks tries to move up to max_load_move weighted load from busiest to
3181 * this_rq, as part of a balancing operation within domain "sd".
3182 * Returns 1 if successful and 0 otherwise.
3184 * Called with both runqueues locked.
3186 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3187 unsigned long max_load_move,
3188 struct sched_domain *sd, enum cpu_idle_type idle,
3191 const struct sched_class *class = sched_class_highest;
3192 unsigned long total_load_moved = 0;
3193 int this_best_prio = this_rq->curr->prio;
3197 class->load_balance(this_rq, this_cpu, busiest,
3198 max_load_move - total_load_moved,
3199 sd, idle, all_pinned, &this_best_prio);
3200 class = class->next;
3201 } while (class && max_load_move > total_load_moved);
3203 return total_load_moved > 0;
3207 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3208 struct sched_domain *sd, enum cpu_idle_type idle,
3209 struct rq_iterator *iterator)
3211 struct task_struct *p = iterator->start(iterator->arg);
3215 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3216 pull_task(busiest, p, this_rq, this_cpu);
3218 * Right now, this is only the second place pull_task()
3219 * is called, so we can safely collect pull_task()
3220 * stats here rather than inside pull_task().
3222 schedstat_inc(sd, lb_gained[idle]);
3226 p = iterator->next(iterator->arg);
3233 * move_one_task tries to move exactly one task from busiest to this_rq, as
3234 * part of active balancing operations within "domain".
3235 * Returns 1 if successful and 0 otherwise.
3237 * Called with both runqueues locked.
3239 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3240 struct sched_domain *sd, enum cpu_idle_type idle)
3242 const struct sched_class *class;
3244 for (class = sched_class_highest; class; class = class->next)
3245 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3252 * find_busiest_group finds and returns the busiest CPU group within the
3253 * domain. It calculates and returns the amount of weighted load which
3254 * should be moved to restore balance via the imbalance parameter.
3256 static struct sched_group *
3257 find_busiest_group(struct sched_domain *sd, int this_cpu,
3258 unsigned long *imbalance, enum cpu_idle_type idle,
3259 int *sd_idle, const cpumask_t *cpus, int *balance)
3261 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3262 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3263 unsigned long max_pull;
3264 unsigned long busiest_load_per_task, busiest_nr_running;
3265 unsigned long this_load_per_task, this_nr_running;
3266 int load_idx, group_imb = 0;
3267 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3268 int power_savings_balance = 1;
3269 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3270 unsigned long min_nr_running = ULONG_MAX;
3271 struct sched_group *group_min = NULL, *group_leader = NULL;
3274 max_load = this_load = total_load = total_pwr = 0;
3275 busiest_load_per_task = busiest_nr_running = 0;
3276 this_load_per_task = this_nr_running = 0;
3277 if (idle == CPU_NOT_IDLE)
3278 load_idx = sd->busy_idx;
3279 else if (idle == CPU_NEWLY_IDLE)
3280 load_idx = sd->newidle_idx;
3282 load_idx = sd->idle_idx;
3285 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3288 int __group_imb = 0;
3289 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3290 unsigned long sum_nr_running, sum_weighted_load;
3292 local_group = cpu_isset(this_cpu, group->cpumask);
3295 balance_cpu = first_cpu(group->cpumask);
3297 /* Tally up the load of all CPUs in the group */
3298 sum_weighted_load = sum_nr_running = avg_load = 0;
3300 min_cpu_load = ~0UL;
3302 for_each_cpu_mask(i, group->cpumask) {
3305 if (!cpu_isset(i, *cpus))
3310 if (*sd_idle && rq->nr_running)
3313 /* Bias balancing toward cpus of our domain */
3315 if (idle_cpu(i) && !first_idle_cpu) {
3320 load = target_load(i, load_idx);
3322 load = source_load(i, load_idx);
3323 if (load > max_cpu_load)
3324 max_cpu_load = load;
3325 if (min_cpu_load > load)
3326 min_cpu_load = load;
3330 sum_nr_running += rq->nr_running;
3331 sum_weighted_load += weighted_cpuload(i);
3335 * First idle cpu or the first cpu(busiest) in this sched group
3336 * is eligible for doing load balancing at this and above
3337 * domains. In the newly idle case, we will allow all the cpu's
3338 * to do the newly idle load balance.
3340 if (idle != CPU_NEWLY_IDLE && local_group &&
3341 balance_cpu != this_cpu && balance) {
3346 total_load += avg_load;
3347 total_pwr += group->__cpu_power;
3349 /* Adjust by relative CPU power of the group */
3350 avg_load = sg_div_cpu_power(group,
3351 avg_load * SCHED_LOAD_SCALE);
3353 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3356 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3359 this_load = avg_load;
3361 this_nr_running = sum_nr_running;
3362 this_load_per_task = sum_weighted_load;
3363 } else if (avg_load > max_load &&
3364 (sum_nr_running > group_capacity || __group_imb)) {
3365 max_load = avg_load;
3367 busiest_nr_running = sum_nr_running;
3368 busiest_load_per_task = sum_weighted_load;
3369 group_imb = __group_imb;
3372 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3374 * Busy processors will not participate in power savings
3377 if (idle == CPU_NOT_IDLE ||
3378 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3382 * If the local group is idle or completely loaded
3383 * no need to do power savings balance at this domain
3385 if (local_group && (this_nr_running >= group_capacity ||
3387 power_savings_balance = 0;
3390 * If a group is already running at full capacity or idle,
3391 * don't include that group in power savings calculations
3393 if (!power_savings_balance || sum_nr_running >= group_capacity
3398 * Calculate the group which has the least non-idle load.
3399 * This is the group from where we need to pick up the load
3402 if ((sum_nr_running < min_nr_running) ||
3403 (sum_nr_running == min_nr_running &&
3404 first_cpu(group->cpumask) <
3405 first_cpu(group_min->cpumask))) {
3407 min_nr_running = sum_nr_running;
3408 min_load_per_task = sum_weighted_load /
3413 * Calculate the group which is almost near its
3414 * capacity but still has some space to pick up some load
3415 * from other group and save more power
3417 if (sum_nr_running <= group_capacity - 1) {
3418 if (sum_nr_running > leader_nr_running ||
3419 (sum_nr_running == leader_nr_running &&
3420 first_cpu(group->cpumask) >
3421 first_cpu(group_leader->cpumask))) {
3422 group_leader = group;
3423 leader_nr_running = sum_nr_running;
3428 group = group->next;
3429 } while (group != sd->groups);
3431 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3434 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3436 if (this_load >= avg_load ||
3437 100*max_load <= sd->imbalance_pct*this_load)
3440 busiest_load_per_task /= busiest_nr_running;
3442 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3445 * We're trying to get all the cpus to the average_load, so we don't
3446 * want to push ourselves above the average load, nor do we wish to
3447 * reduce the max loaded cpu below the average load, as either of these
3448 * actions would just result in more rebalancing later, and ping-pong
3449 * tasks around. Thus we look for the minimum possible imbalance.
3450 * Negative imbalances (*we* are more loaded than anyone else) will
3451 * be counted as no imbalance for these purposes -- we can't fix that
3452 * by pulling tasks to us. Be careful of negative numbers as they'll
3453 * appear as very large values with unsigned longs.
3455 if (max_load <= busiest_load_per_task)
3459 * In the presence of smp nice balancing, certain scenarios can have
3460 * max load less than avg load(as we skip the groups at or below
3461 * its cpu_power, while calculating max_load..)
3463 if (max_load < avg_load) {
3465 goto small_imbalance;
3468 /* Don't want to pull so many tasks that a group would go idle */
3469 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3471 /* How much load to actually move to equalise the imbalance */
3472 *imbalance = min(max_pull * busiest->__cpu_power,
3473 (avg_load - this_load) * this->__cpu_power)
3477 * if *imbalance is less than the average load per runnable task
3478 * there is no gaurantee that any tasks will be moved so we'll have
3479 * a think about bumping its value to force at least one task to be
3482 if (*imbalance < busiest_load_per_task) {
3483 unsigned long tmp, pwr_now, pwr_move;
3487 pwr_move = pwr_now = 0;
3489 if (this_nr_running) {
3490 this_load_per_task /= this_nr_running;
3491 if (busiest_load_per_task > this_load_per_task)
3494 this_load_per_task = SCHED_LOAD_SCALE;
3496 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3497 busiest_load_per_task * imbn) {
3498 *imbalance = busiest_load_per_task;
3503 * OK, we don't have enough imbalance to justify moving tasks,
3504 * however we may be able to increase total CPU power used by
3508 pwr_now += busiest->__cpu_power *
3509 min(busiest_load_per_task, max_load);
3510 pwr_now += this->__cpu_power *
3511 min(this_load_per_task, this_load);
3512 pwr_now /= SCHED_LOAD_SCALE;
3514 /* Amount of load we'd subtract */
3515 tmp = sg_div_cpu_power(busiest,
3516 busiest_load_per_task * SCHED_LOAD_SCALE);
3518 pwr_move += busiest->__cpu_power *
3519 min(busiest_load_per_task, max_load - tmp);
3521 /* Amount of load we'd add */
3522 if (max_load * busiest->__cpu_power <
3523 busiest_load_per_task * SCHED_LOAD_SCALE)
3524 tmp = sg_div_cpu_power(this,
3525 max_load * busiest->__cpu_power);
3527 tmp = sg_div_cpu_power(this,
3528 busiest_load_per_task * SCHED_LOAD_SCALE);
3529 pwr_move += this->__cpu_power *
3530 min(this_load_per_task, this_load + tmp);
3531 pwr_move /= SCHED_LOAD_SCALE;
3533 /* Move if we gain throughput */
3534 if (pwr_move > pwr_now)
3535 *imbalance = busiest_load_per_task;
3541 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3542 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3545 if (this == group_leader && group_leader != group_min) {
3546 *imbalance = min_load_per_task;
3556 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3559 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3560 unsigned long imbalance, const cpumask_t *cpus)
3562 struct rq *busiest = NULL, *rq;
3563 unsigned long max_load = 0;
3566 for_each_cpu_mask(i, group->cpumask) {
3569 if (!cpu_isset(i, *cpus))
3573 wl = weighted_cpuload(i);
3575 if (rq->nr_running == 1 && wl > imbalance)
3578 if (wl > max_load) {
3588 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3589 * so long as it is large enough.
3591 #define MAX_PINNED_INTERVAL 512
3594 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3595 * tasks if there is an imbalance.
3597 static int load_balance(int this_cpu, struct rq *this_rq,
3598 struct sched_domain *sd, enum cpu_idle_type idle,
3599 int *balance, cpumask_t *cpus)
3601 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3602 struct sched_group *group;
3603 unsigned long imbalance;
3605 unsigned long flags;
3606 int unlock_aggregate;
3610 unlock_aggregate = get_aggregate(sd);
3613 * When power savings policy is enabled for the parent domain, idle
3614 * sibling can pick up load irrespective of busy siblings. In this case,
3615 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3616 * portraying it as CPU_NOT_IDLE.
3618 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3619 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3622 schedstat_inc(sd, lb_count[idle]);
3625 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3632 schedstat_inc(sd, lb_nobusyg[idle]);
3636 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3638 schedstat_inc(sd, lb_nobusyq[idle]);
3642 BUG_ON(busiest == this_rq);
3644 schedstat_add(sd, lb_imbalance[idle], imbalance);
3647 if (busiest->nr_running > 1) {
3649 * Attempt to move tasks. If find_busiest_group has found
3650 * an imbalance but busiest->nr_running <= 1, the group is
3651 * still unbalanced. ld_moved simply stays zero, so it is
3652 * correctly treated as an imbalance.
3654 local_irq_save(flags);
3655 double_rq_lock(this_rq, busiest);
3656 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3657 imbalance, sd, idle, &all_pinned);
3658 double_rq_unlock(this_rq, busiest);
3659 local_irq_restore(flags);
3662 * some other cpu did the load balance for us.
3664 if (ld_moved && this_cpu != smp_processor_id())
3665 resched_cpu(this_cpu);
3667 /* All tasks on this runqueue were pinned by CPU affinity */
3668 if (unlikely(all_pinned)) {
3669 cpu_clear(cpu_of(busiest), *cpus);
3670 if (!cpus_empty(*cpus))
3677 schedstat_inc(sd, lb_failed[idle]);
3678 sd->nr_balance_failed++;
3680 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3682 spin_lock_irqsave(&busiest->lock, flags);
3684 /* don't kick the migration_thread, if the curr
3685 * task on busiest cpu can't be moved to this_cpu
3687 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3688 spin_unlock_irqrestore(&busiest->lock, flags);
3690 goto out_one_pinned;
3693 if (!busiest->active_balance) {
3694 busiest->active_balance = 1;
3695 busiest->push_cpu = this_cpu;
3698 spin_unlock_irqrestore(&busiest->lock, flags);
3700 wake_up_process(busiest->migration_thread);
3703 * We've kicked active balancing, reset the failure
3706 sd->nr_balance_failed = sd->cache_nice_tries+1;
3709 sd->nr_balance_failed = 0;
3711 if (likely(!active_balance)) {
3712 /* We were unbalanced, so reset the balancing interval */
3713 sd->balance_interval = sd->min_interval;
3716 * If we've begun active balancing, start to back off. This
3717 * case may not be covered by the all_pinned logic if there
3718 * is only 1 task on the busy runqueue (because we don't call
3721 if (sd->balance_interval < sd->max_interval)
3722 sd->balance_interval *= 2;
3725 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3726 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3732 schedstat_inc(sd, lb_balanced[idle]);
3734 sd->nr_balance_failed = 0;
3737 /* tune up the balancing interval */
3738 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3739 (sd->balance_interval < sd->max_interval))
3740 sd->balance_interval *= 2;
3742 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3743 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3748 if (unlock_aggregate)
3754 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3755 * tasks if there is an imbalance.
3757 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3758 * this_rq is locked.
3761 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3764 struct sched_group *group;
3765 struct rq *busiest = NULL;
3766 unsigned long imbalance;
3774 * When power savings policy is enabled for the parent domain, idle
3775 * sibling can pick up load irrespective of busy siblings. In this case,
3776 * let the state of idle sibling percolate up as IDLE, instead of
3777 * portraying it as CPU_NOT_IDLE.
3779 if (sd->flags & SD_SHARE_CPUPOWER &&
3780 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3783 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3785 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3786 &sd_idle, cpus, NULL);
3788 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3792 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3794 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3798 BUG_ON(busiest == this_rq);
3800 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3803 if (busiest->nr_running > 1) {
3804 /* Attempt to move tasks */
3805 double_lock_balance(this_rq, busiest);
3806 /* this_rq->clock is already updated */
3807 update_rq_clock(busiest);
3808 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3809 imbalance, sd, CPU_NEWLY_IDLE,
3811 spin_unlock(&busiest->lock);
3813 if (unlikely(all_pinned)) {
3814 cpu_clear(cpu_of(busiest), *cpus);
3815 if (!cpus_empty(*cpus))
3821 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3822 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3823 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3826 sd->nr_balance_failed = 0;
3831 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3832 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3833 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3835 sd->nr_balance_failed = 0;
3841 * idle_balance is called by schedule() if this_cpu is about to become
3842 * idle. Attempts to pull tasks from other CPUs.
3844 static void idle_balance(int this_cpu, struct rq *this_rq)
3846 struct sched_domain *sd;
3847 int pulled_task = -1;
3848 unsigned long next_balance = jiffies + HZ;
3851 for_each_domain(this_cpu, sd) {
3852 unsigned long interval;
3854 if (!(sd->flags & SD_LOAD_BALANCE))
3857 if (sd->flags & SD_BALANCE_NEWIDLE)
3858 /* If we've pulled tasks over stop searching: */
3859 pulled_task = load_balance_newidle(this_cpu, this_rq,
3862 interval = msecs_to_jiffies(sd->balance_interval);
3863 if (time_after(next_balance, sd->last_balance + interval))
3864 next_balance = sd->last_balance + interval;
3868 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3870 * We are going idle. next_balance may be set based on
3871 * a busy processor. So reset next_balance.
3873 this_rq->next_balance = next_balance;
3878 * active_load_balance is run by migration threads. It pushes running tasks
3879 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3880 * running on each physical CPU where possible, and avoids physical /
3881 * logical imbalances.
3883 * Called with busiest_rq locked.
3885 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3887 int target_cpu = busiest_rq->push_cpu;
3888 struct sched_domain *sd;
3889 struct rq *target_rq;
3891 /* Is there any task to move? */
3892 if (busiest_rq->nr_running <= 1)
3895 target_rq = cpu_rq(target_cpu);
3898 * This condition is "impossible", if it occurs
3899 * we need to fix it. Originally reported by
3900 * Bjorn Helgaas on a 128-cpu setup.
3902 BUG_ON(busiest_rq == target_rq);
3904 /* move a task from busiest_rq to target_rq */
3905 double_lock_balance(busiest_rq, target_rq);
3906 update_rq_clock(busiest_rq);
3907 update_rq_clock(target_rq);
3909 /* Search for an sd spanning us and the target CPU. */
3910 for_each_domain(target_cpu, sd) {
3911 if ((sd->flags & SD_LOAD_BALANCE) &&
3912 cpu_isset(busiest_cpu, sd->span))
3917 schedstat_inc(sd, alb_count);
3919 if (move_one_task(target_rq, target_cpu, busiest_rq,
3921 schedstat_inc(sd, alb_pushed);
3923 schedstat_inc(sd, alb_failed);
3925 spin_unlock(&target_rq->lock);
3930 atomic_t load_balancer;
3932 } nohz ____cacheline_aligned = {
3933 .load_balancer = ATOMIC_INIT(-1),
3934 .cpu_mask = CPU_MASK_NONE,
3938 * This routine will try to nominate the ilb (idle load balancing)
3939 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3940 * load balancing on behalf of all those cpus. If all the cpus in the system
3941 * go into this tickless mode, then there will be no ilb owner (as there is
3942 * no need for one) and all the cpus will sleep till the next wakeup event
3945 * For the ilb owner, tick is not stopped. And this tick will be used
3946 * for idle load balancing. ilb owner will still be part of
3949 * While stopping the tick, this cpu will become the ilb owner if there
3950 * is no other owner. And will be the owner till that cpu becomes busy
3951 * or if all cpus in the system stop their ticks at which point
3952 * there is no need for ilb owner.
3954 * When the ilb owner becomes busy, it nominates another owner, during the
3955 * next busy scheduler_tick()
3957 int select_nohz_load_balancer(int stop_tick)
3959 int cpu = smp_processor_id();
3962 cpu_set(cpu, nohz.cpu_mask);
3963 cpu_rq(cpu)->in_nohz_recently = 1;
3966 * If we are going offline and still the leader, give up!
3968 if (cpu_is_offline(cpu) &&
3969 atomic_read(&nohz.load_balancer) == cpu) {
3970 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3975 /* time for ilb owner also to sleep */
3976 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3977 if (atomic_read(&nohz.load_balancer) == cpu)
3978 atomic_set(&nohz.load_balancer, -1);
3982 if (atomic_read(&nohz.load_balancer) == -1) {
3983 /* make me the ilb owner */
3984 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3986 } else if (atomic_read(&nohz.load_balancer) == cpu)
3989 if (!cpu_isset(cpu, nohz.cpu_mask))
3992 cpu_clear(cpu, nohz.cpu_mask);
3994 if (atomic_read(&nohz.load_balancer) == cpu)
3995 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4002 static DEFINE_SPINLOCK(balancing);
4005 * It checks each scheduling domain to see if it is due to be balanced,
4006 * and initiates a balancing operation if so.
4008 * Balancing parameters are set up in arch_init_sched_domains.
4010 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4013 struct rq *rq = cpu_rq(cpu);
4014 unsigned long interval;
4015 struct sched_domain *sd;
4016 /* Earliest time when we have to do rebalance again */
4017 unsigned long next_balance = jiffies + 60*HZ;
4018 int update_next_balance = 0;
4021 for_each_domain(cpu, sd) {
4022 if (!(sd->flags & SD_LOAD_BALANCE))
4025 interval = sd->balance_interval;
4026 if (idle != CPU_IDLE)
4027 interval *= sd->busy_factor;
4029 /* scale ms to jiffies */
4030 interval = msecs_to_jiffies(interval);
4031 if (unlikely(!interval))
4033 if (interval > HZ*NR_CPUS/10)
4034 interval = HZ*NR_CPUS/10;
4037 if (sd->flags & SD_SERIALIZE) {
4038 if (!spin_trylock(&balancing))
4042 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4043 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
4045 * We've pulled tasks over so either we're no
4046 * longer idle, or one of our SMT siblings is
4049 idle = CPU_NOT_IDLE;
4051 sd->last_balance = jiffies;
4053 if (sd->flags & SD_SERIALIZE)
4054 spin_unlock(&balancing);
4056 if (time_after(next_balance, sd->last_balance + interval)) {
4057 next_balance = sd->last_balance + interval;
4058 update_next_balance = 1;
4062 * Stop the load balance at this level. There is another
4063 * CPU in our sched group which is doing load balancing more
4071 * next_balance will be updated only when there is a need.
4072 * When the cpu is attached to null domain for ex, it will not be
4075 if (likely(update_next_balance))
4076 rq->next_balance = next_balance;
4080 * run_rebalance_domains is triggered when needed from the scheduler tick.
4081 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4082 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4084 static void run_rebalance_domains(struct softirq_action *h)
4086 int this_cpu = smp_processor_id();
4087 struct rq *this_rq = cpu_rq(this_cpu);
4088 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4089 CPU_IDLE : CPU_NOT_IDLE;
4091 rebalance_domains(this_cpu, idle);
4095 * If this cpu is the owner for idle load balancing, then do the
4096 * balancing on behalf of the other idle cpus whose ticks are
4099 if (this_rq->idle_at_tick &&
4100 atomic_read(&nohz.load_balancer) == this_cpu) {
4101 cpumask_t cpus = nohz.cpu_mask;
4105 cpu_clear(this_cpu, cpus);
4106 for_each_cpu_mask(balance_cpu, cpus) {
4108 * If this cpu gets work to do, stop the load balancing
4109 * work being done for other cpus. Next load
4110 * balancing owner will pick it up.
4115 rebalance_domains(balance_cpu, CPU_IDLE);
4117 rq = cpu_rq(balance_cpu);
4118 if (time_after(this_rq->next_balance, rq->next_balance))
4119 this_rq->next_balance = rq->next_balance;
4126 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4128 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4129 * idle load balancing owner or decide to stop the periodic load balancing,
4130 * if the whole system is idle.
4132 static inline void trigger_load_balance(struct rq *rq, int cpu)
4136 * If we were in the nohz mode recently and busy at the current
4137 * scheduler tick, then check if we need to nominate new idle
4140 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4141 rq->in_nohz_recently = 0;
4143 if (atomic_read(&nohz.load_balancer) == cpu) {
4144 cpu_clear(cpu, nohz.cpu_mask);
4145 atomic_set(&nohz.load_balancer, -1);
4148 if (atomic_read(&nohz.load_balancer) == -1) {
4150 * simple selection for now: Nominate the
4151 * first cpu in the nohz list to be the next
4154 * TBD: Traverse the sched domains and nominate
4155 * the nearest cpu in the nohz.cpu_mask.
4157 int ilb = first_cpu(nohz.cpu_mask);
4159 if (ilb < nr_cpu_ids)
4165 * If this cpu is idle and doing idle load balancing for all the
4166 * cpus with ticks stopped, is it time for that to stop?
4168 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4169 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4175 * If this cpu is idle and the idle load balancing is done by
4176 * someone else, then no need raise the SCHED_SOFTIRQ
4178 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4179 cpu_isset(cpu, nohz.cpu_mask))
4182 if (time_after_eq(jiffies, rq->next_balance))
4183 raise_softirq(SCHED_SOFTIRQ);
4186 #else /* CONFIG_SMP */
4189 * on UP we do not need to balance between CPUs:
4191 static inline void idle_balance(int cpu, struct rq *rq)
4197 DEFINE_PER_CPU(struct kernel_stat, kstat);
4199 EXPORT_PER_CPU_SYMBOL(kstat);
4202 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4203 * that have not yet been banked in case the task is currently running.
4205 unsigned long long task_sched_runtime(struct task_struct *p)
4207 unsigned long flags;
4211 rq = task_rq_lock(p, &flags);
4212 ns = p->se.sum_exec_runtime;
4213 if (task_current(rq, p)) {
4214 update_rq_clock(rq);
4215 delta_exec = rq->clock - p->se.exec_start;
4216 if ((s64)delta_exec > 0)
4219 task_rq_unlock(rq, &flags);
4225 * Account user cpu time to a process.
4226 * @p: the process that the cpu time gets accounted to
4227 * @cputime: the cpu time spent in user space since the last update
4229 void account_user_time(struct task_struct *p, cputime_t cputime)
4231 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4234 p->utime = cputime_add(p->utime, cputime);
4236 /* Add user time to cpustat. */
4237 tmp = cputime_to_cputime64(cputime);
4238 if (TASK_NICE(p) > 0)
4239 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4241 cpustat->user = cputime64_add(cpustat->user, tmp);
4245 * Account guest cpu time to a process.
4246 * @p: the process that the cpu time gets accounted to
4247 * @cputime: the cpu time spent in virtual machine since the last update
4249 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4252 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4254 tmp = cputime_to_cputime64(cputime);
4256 p->utime = cputime_add(p->utime, cputime);
4257 p->gtime = cputime_add(p->gtime, cputime);
4259 cpustat->user = cputime64_add(cpustat->user, tmp);
4260 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4264 * Account scaled user cpu time to a process.
4265 * @p: the process that the cpu time gets accounted to
4266 * @cputime: the cpu time spent in user space since the last update
4268 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4270 p->utimescaled = cputime_add(p->utimescaled, cputime);
4274 * Account system cpu time to a process.
4275 * @p: the process that the cpu time gets accounted to
4276 * @hardirq_offset: the offset to subtract from hardirq_count()
4277 * @cputime: the cpu time spent in kernel space since the last update
4279 void account_system_time(struct task_struct *p, int hardirq_offset,
4282 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4283 struct rq *rq = this_rq();
4286 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4287 account_guest_time(p, cputime);
4291 p->stime = cputime_add(p->stime, cputime);
4293 /* Add system time to cpustat. */
4294 tmp = cputime_to_cputime64(cputime);
4295 if (hardirq_count() - hardirq_offset)
4296 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4297 else if (softirq_count())
4298 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4299 else if (p != rq->idle)
4300 cpustat->system = cputime64_add(cpustat->system, tmp);
4301 else if (atomic_read(&rq->nr_iowait) > 0)
4302 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4304 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4305 /* Account for system time used */
4306 acct_update_integrals(p);
4310 * Account scaled system cpu time to a process.
4311 * @p: the process that the cpu time gets accounted to
4312 * @hardirq_offset: the offset to subtract from hardirq_count()
4313 * @cputime: the cpu time spent in kernel space since the last update
4315 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4317 p->stimescaled = cputime_add(p->stimescaled, cputime);
4321 * Account for involuntary wait time.
4322 * @p: the process from which the cpu time has been stolen
4323 * @steal: the cpu time spent in involuntary wait
4325 void account_steal_time(struct task_struct *p, cputime_t steal)
4327 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4328 cputime64_t tmp = cputime_to_cputime64(steal);
4329 struct rq *rq = this_rq();
4331 if (p == rq->idle) {
4332 p->stime = cputime_add(p->stime, steal);
4333 if (atomic_read(&rq->nr_iowait) > 0)
4334 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4336 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4338 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4342 * This function gets called by the timer code, with HZ frequency.
4343 * We call it with interrupts disabled.
4345 * It also gets called by the fork code, when changing the parent's
4348 void scheduler_tick(void)
4350 int cpu = smp_processor_id();
4351 struct rq *rq = cpu_rq(cpu);
4352 struct task_struct *curr = rq->curr;
4356 spin_lock(&rq->lock);
4357 update_rq_clock(rq);
4358 update_cpu_load(rq);
4359 curr->sched_class->task_tick(rq, curr, 0);
4360 spin_unlock(&rq->lock);
4363 rq->idle_at_tick = idle_cpu(cpu);
4364 trigger_load_balance(rq, cpu);
4368 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4370 void __kprobes add_preempt_count(int val)
4375 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4377 preempt_count() += val;
4379 * Spinlock count overflowing soon?
4381 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4384 EXPORT_SYMBOL(add_preempt_count);
4386 void __kprobes sub_preempt_count(int val)
4391 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4394 * Is the spinlock portion underflowing?
4396 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4397 !(preempt_count() & PREEMPT_MASK)))
4400 preempt_count() -= val;
4402 EXPORT_SYMBOL(sub_preempt_count);
4407 * Print scheduling while atomic bug:
4409 static noinline void __schedule_bug(struct task_struct *prev)
4411 struct pt_regs *regs = get_irq_regs();
4413 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4414 prev->comm, prev->pid, preempt_count());
4416 debug_show_held_locks(prev);
4417 if (irqs_disabled())
4418 print_irqtrace_events(prev);
4427 * Various schedule()-time debugging checks and statistics:
4429 static inline void schedule_debug(struct task_struct *prev)
4432 * Test if we are atomic. Since do_exit() needs to call into
4433 * schedule() atomically, we ignore that path for now.
4434 * Otherwise, whine if we are scheduling when we should not be.
4436 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
4437 __schedule_bug(prev);
4439 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4441 schedstat_inc(this_rq(), sched_count);
4442 #ifdef CONFIG_SCHEDSTATS
4443 if (unlikely(prev->lock_depth >= 0)) {
4444 schedstat_inc(this_rq(), bkl_count);
4445 schedstat_inc(prev, sched_info.bkl_count);
4451 * Pick up the highest-prio task:
4453 static inline struct task_struct *
4454 pick_next_task(struct rq *rq, struct task_struct *prev)
4456 const struct sched_class *class;
4457 struct task_struct *p;
4460 * Optimization: we know that if all tasks are in
4461 * the fair class we can call that function directly:
4463 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4464 p = fair_sched_class.pick_next_task(rq);
4469 class = sched_class_highest;
4471 p = class->pick_next_task(rq);
4475 * Will never be NULL as the idle class always
4476 * returns a non-NULL p:
4478 class = class->next;
4483 * schedule() is the main scheduler function.
4485 asmlinkage void __sched schedule(void)
4487 struct task_struct *prev, *next;
4488 unsigned long *switch_count;
4494 cpu = smp_processor_id();
4498 switch_count = &prev->nivcsw;
4500 release_kernel_lock(prev);
4501 need_resched_nonpreemptible:
4503 schedule_debug(prev);
4508 * Do the rq-clock update outside the rq lock:
4510 local_irq_disable();
4511 update_rq_clock(rq);
4512 spin_lock(&rq->lock);
4513 clear_tsk_need_resched(prev);
4515 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4516 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
4517 signal_pending(prev))) {
4518 prev->state = TASK_RUNNING;
4520 deactivate_task(rq, prev, 1);
4522 switch_count = &prev->nvcsw;
4526 if (prev->sched_class->pre_schedule)
4527 prev->sched_class->pre_schedule(rq, prev);
4530 if (unlikely(!rq->nr_running))
4531 idle_balance(cpu, rq);
4533 prev->sched_class->put_prev_task(rq, prev);
4534 next = pick_next_task(rq, prev);
4536 if (likely(prev != next)) {
4537 sched_info_switch(prev, next);
4543 context_switch(rq, prev, next); /* unlocks the rq */
4545 * the context switch might have flipped the stack from under
4546 * us, hence refresh the local variables.
4548 cpu = smp_processor_id();
4551 spin_unlock_irq(&rq->lock);
4555 if (unlikely(reacquire_kernel_lock(current) < 0))
4556 goto need_resched_nonpreemptible;
4558 preempt_enable_no_resched();
4559 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4562 EXPORT_SYMBOL(schedule);
4564 #ifdef CONFIG_PREEMPT
4566 * this is the entry point to schedule() from in-kernel preemption
4567 * off of preempt_enable. Kernel preemptions off return from interrupt
4568 * occur there and call schedule directly.
4570 asmlinkage void __sched preempt_schedule(void)
4572 struct thread_info *ti = current_thread_info();
4575 * If there is a non-zero preempt_count or interrupts are disabled,
4576 * we do not want to preempt the current task. Just return..
4578 if (likely(ti->preempt_count || irqs_disabled()))
4582 add_preempt_count(PREEMPT_ACTIVE);
4584 sub_preempt_count(PREEMPT_ACTIVE);
4587 * Check again in case we missed a preemption opportunity
4588 * between schedule and now.
4591 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4593 EXPORT_SYMBOL(preempt_schedule);
4596 * this is the entry point to schedule() from kernel preemption
4597 * off of irq context.
4598 * Note, that this is called and return with irqs disabled. This will
4599 * protect us against recursive calling from irq.
4601 asmlinkage void __sched preempt_schedule_irq(void)
4603 struct thread_info *ti = current_thread_info();
4605 /* Catch callers which need to be fixed */
4606 BUG_ON(ti->preempt_count || !irqs_disabled());
4609 add_preempt_count(PREEMPT_ACTIVE);
4612 local_irq_disable();
4613 sub_preempt_count(PREEMPT_ACTIVE);
4616 * Check again in case we missed a preemption opportunity
4617 * between schedule and now.
4620 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4623 #endif /* CONFIG_PREEMPT */
4625 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4628 return try_to_wake_up(curr->private, mode, sync);
4630 EXPORT_SYMBOL(default_wake_function);
4633 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4634 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4635 * number) then we wake all the non-exclusive tasks and one exclusive task.
4637 * There are circumstances in which we can try to wake a task which has already
4638 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4639 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4641 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4642 int nr_exclusive, int sync, void *key)
4644 wait_queue_t *curr, *next;
4646 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4647 unsigned flags = curr->flags;
4649 if (curr->func(curr, mode, sync, key) &&
4650 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4656 * __wake_up - wake up threads blocked on a waitqueue.
4658 * @mode: which threads
4659 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4660 * @key: is directly passed to the wakeup function
4662 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4663 int nr_exclusive, void *key)
4665 unsigned long flags;
4667 spin_lock_irqsave(&q->lock, flags);
4668 __wake_up_common(q, mode, nr_exclusive, 0, key);
4669 spin_unlock_irqrestore(&q->lock, flags);
4671 EXPORT_SYMBOL(__wake_up);
4674 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4676 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4678 __wake_up_common(q, mode, 1, 0, NULL);
4682 * __wake_up_sync - wake up threads blocked on a waitqueue.
4684 * @mode: which threads
4685 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4687 * The sync wakeup differs that the waker knows that it will schedule
4688 * away soon, so while the target thread will be woken up, it will not
4689 * be migrated to another CPU - ie. the two threads are 'synchronized'
4690 * with each other. This can prevent needless bouncing between CPUs.
4692 * On UP it can prevent extra preemption.
4695 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4697 unsigned long flags;
4703 if (unlikely(!nr_exclusive))
4706 spin_lock_irqsave(&q->lock, flags);
4707 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4708 spin_unlock_irqrestore(&q->lock, flags);
4710 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4712 void complete(struct completion *x)
4714 unsigned long flags;
4716 spin_lock_irqsave(&x->wait.lock, flags);
4718 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4719 spin_unlock_irqrestore(&x->wait.lock, flags);
4721 EXPORT_SYMBOL(complete);
4723 void complete_all(struct completion *x)
4725 unsigned long flags;
4727 spin_lock_irqsave(&x->wait.lock, flags);
4728 x->done += UINT_MAX/2;
4729 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4730 spin_unlock_irqrestore(&x->wait.lock, flags);
4732 EXPORT_SYMBOL(complete_all);
4734 static inline long __sched
4735 do_wait_for_common(struct completion *x, long timeout, int state)
4738 DECLARE_WAITQUEUE(wait, current);
4740 wait.flags |= WQ_FLAG_EXCLUSIVE;
4741 __add_wait_queue_tail(&x->wait, &wait);
4743 if ((state == TASK_INTERRUPTIBLE &&
4744 signal_pending(current)) ||
4745 (state == TASK_KILLABLE &&
4746 fatal_signal_pending(current))) {
4747 __remove_wait_queue(&x->wait, &wait);
4748 return -ERESTARTSYS;
4750 __set_current_state(state);
4751 spin_unlock_irq(&x->wait.lock);
4752 timeout = schedule_timeout(timeout);
4753 spin_lock_irq(&x->wait.lock);
4755 __remove_wait_queue(&x->wait, &wait);
4759 __remove_wait_queue(&x->wait, &wait);
4766 wait_for_common(struct completion *x, long timeout, int state)
4770 spin_lock_irq(&x->wait.lock);
4771 timeout = do_wait_for_common(x, timeout, state);
4772 spin_unlock_irq(&x->wait.lock);
4776 void __sched wait_for_completion(struct completion *x)
4778 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4780 EXPORT_SYMBOL(wait_for_completion);
4782 unsigned long __sched
4783 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4785 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4787 EXPORT_SYMBOL(wait_for_completion_timeout);
4789 int __sched wait_for_completion_interruptible(struct completion *x)
4791 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4792 if (t == -ERESTARTSYS)
4796 EXPORT_SYMBOL(wait_for_completion_interruptible);
4798 unsigned long __sched
4799 wait_for_completion_interruptible_timeout(struct completion *x,
4800 unsigned long timeout)
4802 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4804 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4806 int __sched wait_for_completion_killable(struct completion *x)
4808 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4809 if (t == -ERESTARTSYS)
4813 EXPORT_SYMBOL(wait_for_completion_killable);
4816 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4818 unsigned long flags;
4821 init_waitqueue_entry(&wait, current);
4823 __set_current_state(state);
4825 spin_lock_irqsave(&q->lock, flags);
4826 __add_wait_queue(q, &wait);
4827 spin_unlock(&q->lock);
4828 timeout = schedule_timeout(timeout);
4829 spin_lock_irq(&q->lock);
4830 __remove_wait_queue(q, &wait);
4831 spin_unlock_irqrestore(&q->lock, flags);
4836 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4838 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4840 EXPORT_SYMBOL(interruptible_sleep_on);
4843 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4845 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4847 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4849 void __sched sleep_on(wait_queue_head_t *q)
4851 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4853 EXPORT_SYMBOL(sleep_on);
4855 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4857 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4859 EXPORT_SYMBOL(sleep_on_timeout);
4861 #ifdef CONFIG_RT_MUTEXES
4864 * rt_mutex_setprio - set the current priority of a task
4866 * @prio: prio value (kernel-internal form)
4868 * This function changes the 'effective' priority of a task. It does
4869 * not touch ->normal_prio like __setscheduler().
4871 * Used by the rt_mutex code to implement priority inheritance logic.
4873 void rt_mutex_setprio(struct task_struct *p, int prio)
4875 unsigned long flags;
4876 int oldprio, on_rq, running;
4878 const struct sched_class *prev_class = p->sched_class;
4880 BUG_ON(prio < 0 || prio > MAX_PRIO);
4882 rq = task_rq_lock(p, &flags);
4883 update_rq_clock(rq);
4886 on_rq = p->se.on_rq;
4887 running = task_current(rq, p);
4889 dequeue_task(rq, p, 0);
4891 p->sched_class->put_prev_task(rq, p);
4894 p->sched_class = &rt_sched_class;
4896 p->sched_class = &fair_sched_class;
4901 p->sched_class->set_curr_task(rq);
4903 enqueue_task(rq, p, 0);
4905 check_class_changed(rq, p, prev_class, oldprio, running);
4907 task_rq_unlock(rq, &flags);
4912 void set_user_nice(struct task_struct *p, long nice)
4914 int old_prio, delta, on_rq;
4915 unsigned long flags;
4918 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4921 * We have to be careful, if called from sys_setpriority(),
4922 * the task might be in the middle of scheduling on another CPU.
4924 rq = task_rq_lock(p, &flags);
4925 update_rq_clock(rq);
4927 * The RT priorities are set via sched_setscheduler(), but we still
4928 * allow the 'normal' nice value to be set - but as expected
4929 * it wont have any effect on scheduling until the task is
4930 * SCHED_FIFO/SCHED_RR:
4932 if (task_has_rt_policy(p)) {
4933 p->static_prio = NICE_TO_PRIO(nice);
4936 on_rq = p->se.on_rq;
4938 dequeue_task(rq, p, 0);
4940 p->static_prio = NICE_TO_PRIO(nice);
4943 p->prio = effective_prio(p);
4944 delta = p->prio - old_prio;
4947 enqueue_task(rq, p, 0);
4949 * If the task increased its priority or is running and
4950 * lowered its priority, then reschedule its CPU:
4952 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4953 resched_task(rq->curr);
4956 task_rq_unlock(rq, &flags);
4958 EXPORT_SYMBOL(set_user_nice);
4961 * can_nice - check if a task can reduce its nice value
4965 int can_nice(const struct task_struct *p, const int nice)
4967 /* convert nice value [19,-20] to rlimit style value [1,40] */
4968 int nice_rlim = 20 - nice;
4970 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4971 capable(CAP_SYS_NICE));
4974 #ifdef __ARCH_WANT_SYS_NICE
4977 * sys_nice - change the priority of the current process.
4978 * @increment: priority increment
4980 * sys_setpriority is a more generic, but much slower function that
4981 * does similar things.
4983 asmlinkage long sys_nice(int increment)
4988 * Setpriority might change our priority at the same moment.
4989 * We don't have to worry. Conceptually one call occurs first
4990 * and we have a single winner.
4992 if (increment < -40)
4997 nice = PRIO_TO_NICE(current->static_prio) + increment;
5003 if (increment < 0 && !can_nice(current, nice))
5006 retval = security_task_setnice(current, nice);
5010 set_user_nice(current, nice);
5017 * task_prio - return the priority value of a given task.
5018 * @p: the task in question.
5020 * This is the priority value as seen by users in /proc.
5021 * RT tasks are offset by -200. Normal tasks are centered
5022 * around 0, value goes from -16 to +15.
5024 int task_prio(const struct task_struct *p)
5026 return p->prio - MAX_RT_PRIO;
5030 * task_nice - return the nice value of a given task.
5031 * @p: the task in question.
5033 int task_nice(const struct task_struct *p)
5035 return TASK_NICE(p);
5037 EXPORT_SYMBOL(task_nice);
5040 * idle_cpu - is a given cpu idle currently?
5041 * @cpu: the processor in question.
5043 int idle_cpu(int cpu)
5045 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5049 * idle_task - return the idle task for a given cpu.
5050 * @cpu: the processor in question.
5052 struct task_struct *idle_task(int cpu)
5054 return cpu_rq(cpu)->idle;
5058 * find_process_by_pid - find a process with a matching PID value.
5059 * @pid: the pid in question.
5061 static struct task_struct *find_process_by_pid(pid_t pid)
5063 return pid ? find_task_by_vpid(pid) : current;
5066 /* Actually do priority change: must hold rq lock. */
5068 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5070 BUG_ON(p->se.on_rq);
5073 switch (p->policy) {
5077 p->sched_class = &fair_sched_class;
5081 p->sched_class = &rt_sched_class;
5085 p->rt_priority = prio;
5086 p->normal_prio = normal_prio(p);
5087 /* we are holding p->pi_lock already */
5088 p->prio = rt_mutex_getprio(p);
5093 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5094 * @p: the task in question.
5095 * @policy: new policy.
5096 * @param: structure containing the new RT priority.
5098 * NOTE that the task may be already dead.
5100 int sched_setscheduler(struct task_struct *p, int policy,
5101 struct sched_param *param)
5103 int retval, oldprio, oldpolicy = -1, on_rq, running;
5104 unsigned long flags;
5105 const struct sched_class *prev_class = p->sched_class;
5108 /* may grab non-irq protected spin_locks */
5109 BUG_ON(in_interrupt());
5111 /* double check policy once rq lock held */
5113 policy = oldpolicy = p->policy;
5114 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5115 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5116 policy != SCHED_IDLE)
5119 * Valid priorities for SCHED_FIFO and SCHED_RR are
5120 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5121 * SCHED_BATCH and SCHED_IDLE is 0.
5123 if (param->sched_priority < 0 ||
5124 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5125 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5127 if (rt_policy(policy) != (param->sched_priority != 0))
5131 * Allow unprivileged RT tasks to decrease priority:
5133 if (!capable(CAP_SYS_NICE)) {
5134 if (rt_policy(policy)) {
5135 unsigned long rlim_rtprio;
5137 if (!lock_task_sighand(p, &flags))
5139 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5140 unlock_task_sighand(p, &flags);
5142 /* can't set/change the rt policy */
5143 if (policy != p->policy && !rlim_rtprio)
5146 /* can't increase priority */
5147 if (param->sched_priority > p->rt_priority &&
5148 param->sched_priority > rlim_rtprio)
5152 * Like positive nice levels, dont allow tasks to
5153 * move out of SCHED_IDLE either:
5155 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5158 /* can't change other user's priorities */
5159 if ((current->euid != p->euid) &&
5160 (current->euid != p->uid))
5164 #ifdef CONFIG_RT_GROUP_SCHED
5166 * Do not allow realtime tasks into groups that have no runtime
5169 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
5173 retval = security_task_setscheduler(p, policy, param);
5177 * make sure no PI-waiters arrive (or leave) while we are
5178 * changing the priority of the task:
5180 spin_lock_irqsave(&p->pi_lock, flags);
5182 * To be able to change p->policy safely, the apropriate
5183 * runqueue lock must be held.
5185 rq = __task_rq_lock(p);
5186 /* recheck policy now with rq lock held */
5187 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5188 policy = oldpolicy = -1;
5189 __task_rq_unlock(rq);
5190 spin_unlock_irqrestore(&p->pi_lock, flags);
5193 update_rq_clock(rq);
5194 on_rq = p->se.on_rq;
5195 running = task_current(rq, p);
5197 deactivate_task(rq, p, 0);
5199 p->sched_class->put_prev_task(rq, p);
5202 __setscheduler(rq, p, policy, param->sched_priority);
5205 p->sched_class->set_curr_task(rq);
5207 activate_task(rq, p, 0);
5209 check_class_changed(rq, p, prev_class, oldprio, running);
5211 __task_rq_unlock(rq);
5212 spin_unlock_irqrestore(&p->pi_lock, flags);
5214 rt_mutex_adjust_pi(p);
5218 EXPORT_SYMBOL_GPL(sched_setscheduler);
5221 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5223 struct sched_param lparam;
5224 struct task_struct *p;
5227 if (!param || pid < 0)
5229 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5234 p = find_process_by_pid(pid);
5236 retval = sched_setscheduler(p, policy, &lparam);
5243 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5244 * @pid: the pid in question.
5245 * @policy: new policy.
5246 * @param: structure containing the new RT priority.
5249 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5251 /* negative values for policy are not valid */
5255 return do_sched_setscheduler(pid, policy, param);
5259 * sys_sched_setparam - set/change the RT priority of a thread
5260 * @pid: the pid in question.
5261 * @param: structure containing the new RT priority.
5263 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5265 return do_sched_setscheduler(pid, -1, param);
5269 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5270 * @pid: the pid in question.
5272 asmlinkage long sys_sched_getscheduler(pid_t pid)
5274 struct task_struct *p;
5281 read_lock(&tasklist_lock);
5282 p = find_process_by_pid(pid);
5284 retval = security_task_getscheduler(p);
5288 read_unlock(&tasklist_lock);
5293 * sys_sched_getscheduler - get the RT priority of a thread
5294 * @pid: the pid in question.
5295 * @param: structure containing the RT priority.
5297 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5299 struct sched_param lp;
5300 struct task_struct *p;
5303 if (!param || pid < 0)
5306 read_lock(&tasklist_lock);
5307 p = find_process_by_pid(pid);
5312 retval = security_task_getscheduler(p);
5316 lp.sched_priority = p->rt_priority;
5317 read_unlock(&tasklist_lock);
5320 * This one might sleep, we cannot do it with a spinlock held ...
5322 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5327 read_unlock(&tasklist_lock);
5331 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5333 cpumask_t cpus_allowed;
5334 cpumask_t new_mask = *in_mask;
5335 struct task_struct *p;
5339 read_lock(&tasklist_lock);
5341 p = find_process_by_pid(pid);
5343 read_unlock(&tasklist_lock);
5349 * It is not safe to call set_cpus_allowed with the
5350 * tasklist_lock held. We will bump the task_struct's
5351 * usage count and then drop tasklist_lock.
5354 read_unlock(&tasklist_lock);
5357 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5358 !capable(CAP_SYS_NICE))
5361 retval = security_task_setscheduler(p, 0, NULL);
5365 cpuset_cpus_allowed(p, &cpus_allowed);
5366 cpus_and(new_mask, new_mask, cpus_allowed);
5368 retval = set_cpus_allowed_ptr(p, &new_mask);
5371 cpuset_cpus_allowed(p, &cpus_allowed);
5372 if (!cpus_subset(new_mask, cpus_allowed)) {
5374 * We must have raced with a concurrent cpuset
5375 * update. Just reset the cpus_allowed to the
5376 * cpuset's cpus_allowed
5378 new_mask = cpus_allowed;
5388 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5389 cpumask_t *new_mask)
5391 if (len < sizeof(cpumask_t)) {
5392 memset(new_mask, 0, sizeof(cpumask_t));
5393 } else if (len > sizeof(cpumask_t)) {
5394 len = sizeof(cpumask_t);
5396 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5400 * sys_sched_setaffinity - set the cpu affinity of a process
5401 * @pid: pid of the process
5402 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5403 * @user_mask_ptr: user-space pointer to the new cpu mask
5405 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5406 unsigned long __user *user_mask_ptr)
5411 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5415 return sched_setaffinity(pid, &new_mask);
5419 * Represents all cpu's present in the system
5420 * In systems capable of hotplug, this map could dynamically grow
5421 * as new cpu's are detected in the system via any platform specific
5422 * method, such as ACPI for e.g.
5425 cpumask_t cpu_present_map __read_mostly;
5426 EXPORT_SYMBOL(cpu_present_map);
5429 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
5430 EXPORT_SYMBOL(cpu_online_map);
5432 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
5433 EXPORT_SYMBOL(cpu_possible_map);
5436 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5438 struct task_struct *p;
5442 read_lock(&tasklist_lock);
5445 p = find_process_by_pid(pid);
5449 retval = security_task_getscheduler(p);
5453 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5456 read_unlock(&tasklist_lock);
5463 * sys_sched_getaffinity - get the cpu affinity of a process
5464 * @pid: pid of the process
5465 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5466 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5468 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5469 unsigned long __user *user_mask_ptr)
5474 if (len < sizeof(cpumask_t))
5477 ret = sched_getaffinity(pid, &mask);
5481 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5484 return sizeof(cpumask_t);
5488 * sys_sched_yield - yield the current processor to other threads.
5490 * This function yields the current CPU to other tasks. If there are no
5491 * other threads running on this CPU then this function will return.
5493 asmlinkage long sys_sched_yield(void)
5495 struct rq *rq = this_rq_lock();
5497 schedstat_inc(rq, yld_count);
5498 current->sched_class->yield_task(rq);
5501 * Since we are going to call schedule() anyway, there's
5502 * no need to preempt or enable interrupts:
5504 __release(rq->lock);
5505 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5506 _raw_spin_unlock(&rq->lock);
5507 preempt_enable_no_resched();
5514 static void __cond_resched(void)
5516 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5517 __might_sleep(__FILE__, __LINE__);
5520 * The BKS might be reacquired before we have dropped
5521 * PREEMPT_ACTIVE, which could trigger a second
5522 * cond_resched() call.
5525 add_preempt_count(PREEMPT_ACTIVE);
5527 sub_preempt_count(PREEMPT_ACTIVE);
5528 } while (need_resched());
5531 int __sched _cond_resched(void)
5533 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5534 system_state == SYSTEM_RUNNING) {
5540 EXPORT_SYMBOL(_cond_resched);
5543 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5544 * call schedule, and on return reacquire the lock.
5546 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5547 * operations here to prevent schedule() from being called twice (once via
5548 * spin_unlock(), once by hand).
5550 int cond_resched_lock(spinlock_t *lock)
5552 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5555 if (spin_needbreak(lock) || resched) {
5557 if (resched && need_resched())
5566 EXPORT_SYMBOL(cond_resched_lock);
5568 int __sched cond_resched_softirq(void)
5570 BUG_ON(!in_softirq());
5572 if (need_resched() && system_state == SYSTEM_RUNNING) {
5580 EXPORT_SYMBOL(cond_resched_softirq);
5583 * yield - yield the current processor to other threads.
5585 * This is a shortcut for kernel-space yielding - it marks the
5586 * thread runnable and calls sys_sched_yield().
5588 void __sched yield(void)
5590 set_current_state(TASK_RUNNING);
5593 EXPORT_SYMBOL(yield);
5596 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5597 * that process accounting knows that this is a task in IO wait state.
5599 * But don't do that if it is a deliberate, throttling IO wait (this task
5600 * has set its backing_dev_info: the queue against which it should throttle)
5602 void __sched io_schedule(void)
5604 struct rq *rq = &__raw_get_cpu_var(runqueues);
5606 delayacct_blkio_start();
5607 atomic_inc(&rq->nr_iowait);
5609 atomic_dec(&rq->nr_iowait);
5610 delayacct_blkio_end();
5612 EXPORT_SYMBOL(io_schedule);
5614 long __sched io_schedule_timeout(long timeout)
5616 struct rq *rq = &__raw_get_cpu_var(runqueues);
5619 delayacct_blkio_start();
5620 atomic_inc(&rq->nr_iowait);
5621 ret = schedule_timeout(timeout);
5622 atomic_dec(&rq->nr_iowait);
5623 delayacct_blkio_end();
5628 * sys_sched_get_priority_max - return maximum RT priority.
5629 * @policy: scheduling class.
5631 * this syscall returns the maximum rt_priority that can be used
5632 * by a given scheduling class.
5634 asmlinkage long sys_sched_get_priority_max(int policy)
5641 ret = MAX_USER_RT_PRIO-1;
5653 * sys_sched_get_priority_min - return minimum RT priority.
5654 * @policy: scheduling class.
5656 * this syscall returns the minimum rt_priority that can be used
5657 * by a given scheduling class.
5659 asmlinkage long sys_sched_get_priority_min(int policy)
5677 * sys_sched_rr_get_interval - return the default timeslice of a process.
5678 * @pid: pid of the process.
5679 * @interval: userspace pointer to the timeslice value.
5681 * this syscall writes the default timeslice value of a given process
5682 * into the user-space timespec buffer. A value of '0' means infinity.
5685 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5687 struct task_struct *p;
5688 unsigned int time_slice;
5696 read_lock(&tasklist_lock);
5697 p = find_process_by_pid(pid);
5701 retval = security_task_getscheduler(p);
5706 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5707 * tasks that are on an otherwise idle runqueue:
5710 if (p->policy == SCHED_RR) {
5711 time_slice = DEF_TIMESLICE;
5712 } else if (p->policy != SCHED_FIFO) {
5713 struct sched_entity *se = &p->se;
5714 unsigned long flags;
5717 rq = task_rq_lock(p, &flags);
5718 if (rq->cfs.load.weight)
5719 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5720 task_rq_unlock(rq, &flags);
5722 read_unlock(&tasklist_lock);
5723 jiffies_to_timespec(time_slice, &t);
5724 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5728 read_unlock(&tasklist_lock);
5732 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5734 void sched_show_task(struct task_struct *p)
5736 unsigned long free = 0;
5739 state = p->state ? __ffs(p->state) + 1 : 0;
5740 printk(KERN_INFO "%-13.13s %c", p->comm,
5741 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5742 #if BITS_PER_LONG == 32
5743 if (state == TASK_RUNNING)
5744 printk(KERN_CONT " running ");
5746 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5748 if (state == TASK_RUNNING)
5749 printk(KERN_CONT " running task ");
5751 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5753 #ifdef CONFIG_DEBUG_STACK_USAGE
5755 unsigned long *n = end_of_stack(p);
5758 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5761 printk(KERN_CONT "%5lu %5d %6d\n", free,
5762 task_pid_nr(p), task_pid_nr(p->real_parent));
5764 show_stack(p, NULL);
5767 void show_state_filter(unsigned long state_filter)
5769 struct task_struct *g, *p;
5771 #if BITS_PER_LONG == 32
5773 " task PC stack pid father\n");
5776 " task PC stack pid father\n");
5778 read_lock(&tasklist_lock);
5779 do_each_thread(g, p) {
5781 * reset the NMI-timeout, listing all files on a slow
5782 * console might take alot of time:
5784 touch_nmi_watchdog();
5785 if (!state_filter || (p->state & state_filter))
5787 } while_each_thread(g, p);
5789 touch_all_softlockup_watchdogs();
5791 #ifdef CONFIG_SCHED_DEBUG
5792 sysrq_sched_debug_show();
5794 read_unlock(&tasklist_lock);
5796 * Only show locks if all tasks are dumped:
5798 if (state_filter == -1)
5799 debug_show_all_locks();
5802 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5804 idle->sched_class = &idle_sched_class;
5808 * init_idle - set up an idle thread for a given CPU
5809 * @idle: task in question
5810 * @cpu: cpu the idle task belongs to
5812 * NOTE: this function does not set the idle thread's NEED_RESCHED
5813 * flag, to make booting more robust.
5815 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5817 struct rq *rq = cpu_rq(cpu);
5818 unsigned long flags;
5821 idle->se.exec_start = sched_clock();
5823 idle->prio = idle->normal_prio = MAX_PRIO;
5824 idle->cpus_allowed = cpumask_of_cpu(cpu);
5825 __set_task_cpu(idle, cpu);
5827 spin_lock_irqsave(&rq->lock, flags);
5828 rq->curr = rq->idle = idle;
5829 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5832 spin_unlock_irqrestore(&rq->lock, flags);
5834 /* Set the preempt count _outside_ the spinlocks! */
5835 #if defined(CONFIG_PREEMPT)
5836 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5838 task_thread_info(idle)->preempt_count = 0;
5841 * The idle tasks have their own, simple scheduling class:
5843 idle->sched_class = &idle_sched_class;
5847 * In a system that switches off the HZ timer nohz_cpu_mask
5848 * indicates which cpus entered this state. This is used
5849 * in the rcu update to wait only for active cpus. For system
5850 * which do not switch off the HZ timer nohz_cpu_mask should
5851 * always be CPU_MASK_NONE.
5853 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5856 * Increase the granularity value when there are more CPUs,
5857 * because with more CPUs the 'effective latency' as visible
5858 * to users decreases. But the relationship is not linear,
5859 * so pick a second-best guess by going with the log2 of the
5862 * This idea comes from the SD scheduler of Con Kolivas:
5864 static inline void sched_init_granularity(void)
5866 unsigned int factor = 1 + ilog2(num_online_cpus());
5867 const unsigned long limit = 200000000;
5869 sysctl_sched_min_granularity *= factor;
5870 if (sysctl_sched_min_granularity > limit)
5871 sysctl_sched_min_granularity = limit;
5873 sysctl_sched_latency *= factor;
5874 if (sysctl_sched_latency > limit)
5875 sysctl_sched_latency = limit;
5877 sysctl_sched_wakeup_granularity *= factor;
5882 * This is how migration works:
5884 * 1) we queue a struct migration_req structure in the source CPU's
5885 * runqueue and wake up that CPU's migration thread.
5886 * 2) we down() the locked semaphore => thread blocks.
5887 * 3) migration thread wakes up (implicitly it forces the migrated
5888 * thread off the CPU)
5889 * 4) it gets the migration request and checks whether the migrated
5890 * task is still in the wrong runqueue.
5891 * 5) if it's in the wrong runqueue then the migration thread removes
5892 * it and puts it into the right queue.
5893 * 6) migration thread up()s the semaphore.
5894 * 7) we wake up and the migration is done.
5898 * Change a given task's CPU affinity. Migrate the thread to a
5899 * proper CPU and schedule it away if the CPU it's executing on
5900 * is removed from the allowed bitmask.
5902 * NOTE: the caller must have a valid reference to the task, the
5903 * task must not exit() & deallocate itself prematurely. The
5904 * call is not atomic; no spinlocks may be held.
5906 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5908 struct migration_req req;
5909 unsigned long flags;
5913 rq = task_rq_lock(p, &flags);
5914 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5919 if (p->sched_class->set_cpus_allowed)
5920 p->sched_class->set_cpus_allowed(p, new_mask);
5922 p->cpus_allowed = *new_mask;
5923 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5926 /* Can the task run on the task's current CPU? If so, we're done */
5927 if (cpu_isset(task_cpu(p), *new_mask))
5930 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5931 /* Need help from migration thread: drop lock and wait. */
5932 task_rq_unlock(rq, &flags);
5933 wake_up_process(rq->migration_thread);
5934 wait_for_completion(&req.done);
5935 tlb_migrate_finish(p->mm);
5939 task_rq_unlock(rq, &flags);
5943 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5946 * Move (not current) task off this cpu, onto dest cpu. We're doing
5947 * this because either it can't run here any more (set_cpus_allowed()
5948 * away from this CPU, or CPU going down), or because we're
5949 * attempting to rebalance this task on exec (sched_exec).
5951 * So we race with normal scheduler movements, but that's OK, as long
5952 * as the task is no longer on this CPU.
5954 * Returns non-zero if task was successfully migrated.
5956 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5958 struct rq *rq_dest, *rq_src;
5961 if (unlikely(cpu_is_offline(dest_cpu)))
5964 rq_src = cpu_rq(src_cpu);
5965 rq_dest = cpu_rq(dest_cpu);
5967 double_rq_lock(rq_src, rq_dest);
5968 /* Already moved. */
5969 if (task_cpu(p) != src_cpu)
5971 /* Affinity changed (again). */
5972 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5975 on_rq = p->se.on_rq;
5977 deactivate_task(rq_src, p, 0);
5979 set_task_cpu(p, dest_cpu);
5981 activate_task(rq_dest, p, 0);
5982 check_preempt_curr(rq_dest, p);
5986 double_rq_unlock(rq_src, rq_dest);
5991 * migration_thread - this is a highprio system thread that performs
5992 * thread migration by bumping thread off CPU then 'pushing' onto
5995 static int migration_thread(void *data)
5997 int cpu = (long)data;
6001 BUG_ON(rq->migration_thread != current);
6003 set_current_state(TASK_INTERRUPTIBLE);
6004 while (!kthread_should_stop()) {
6005 struct migration_req *req;
6006 struct list_head *head;
6008 spin_lock_irq(&rq->lock);
6010 if (cpu_is_offline(cpu)) {
6011 spin_unlock_irq(&rq->lock);
6015 if (rq->active_balance) {
6016 active_load_balance(rq, cpu);
6017 rq->active_balance = 0;
6020 head = &rq->migration_queue;
6022 if (list_empty(head)) {
6023 spin_unlock_irq(&rq->lock);
6025 set_current_state(TASK_INTERRUPTIBLE);
6028 req = list_entry(head->next, struct migration_req, list);
6029 list_del_init(head->next);
6031 spin_unlock(&rq->lock);
6032 __migrate_task(req->task, cpu, req->dest_cpu);
6035 complete(&req->done);
6037 __set_current_state(TASK_RUNNING);
6041 /* Wait for kthread_stop */
6042 set_current_state(TASK_INTERRUPTIBLE);
6043 while (!kthread_should_stop()) {
6045 set_current_state(TASK_INTERRUPTIBLE);
6047 __set_current_state(TASK_RUNNING);
6051 #ifdef CONFIG_HOTPLUG_CPU
6053 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6057 local_irq_disable();
6058 ret = __migrate_task(p, src_cpu, dest_cpu);
6064 * Figure out where task on dead CPU should go, use force if necessary.
6065 * NOTE: interrupts should be disabled by the caller
6067 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6069 unsigned long flags;
6076 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6077 cpus_and(mask, mask, p->cpus_allowed);
6078 dest_cpu = any_online_cpu(mask);
6080 /* On any allowed CPU? */
6081 if (dest_cpu >= nr_cpu_ids)
6082 dest_cpu = any_online_cpu(p->cpus_allowed);
6084 /* No more Mr. Nice Guy. */
6085 if (dest_cpu >= nr_cpu_ids) {
6086 cpumask_t cpus_allowed;
6088 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6090 * Try to stay on the same cpuset, where the
6091 * current cpuset may be a subset of all cpus.
6092 * The cpuset_cpus_allowed_locked() variant of
6093 * cpuset_cpus_allowed() will not block. It must be
6094 * called within calls to cpuset_lock/cpuset_unlock.
6096 rq = task_rq_lock(p, &flags);
6097 p->cpus_allowed = cpus_allowed;
6098 dest_cpu = any_online_cpu(p->cpus_allowed);
6099 task_rq_unlock(rq, &flags);
6102 * Don't tell them about moving exiting tasks or
6103 * kernel threads (both mm NULL), since they never
6106 if (p->mm && printk_ratelimit()) {
6107 printk(KERN_INFO "process %d (%s) no "
6108 "longer affine to cpu%d\n",
6109 task_pid_nr(p), p->comm, dead_cpu);
6112 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6116 * While a dead CPU has no uninterruptible tasks queued at this point,
6117 * it might still have a nonzero ->nr_uninterruptible counter, because
6118 * for performance reasons the counter is not stricly tracking tasks to
6119 * their home CPUs. So we just add the counter to another CPU's counter,
6120 * to keep the global sum constant after CPU-down:
6122 static void migrate_nr_uninterruptible(struct rq *rq_src)
6124 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6125 unsigned long flags;
6127 local_irq_save(flags);
6128 double_rq_lock(rq_src, rq_dest);
6129 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6130 rq_src->nr_uninterruptible = 0;
6131 double_rq_unlock(rq_src, rq_dest);
6132 local_irq_restore(flags);
6135 /* Run through task list and migrate tasks from the dead cpu. */
6136 static void migrate_live_tasks(int src_cpu)
6138 struct task_struct *p, *t;
6140 read_lock(&tasklist_lock);
6142 do_each_thread(t, p) {
6146 if (task_cpu(p) == src_cpu)
6147 move_task_off_dead_cpu(src_cpu, p);
6148 } while_each_thread(t, p);
6150 read_unlock(&tasklist_lock);
6154 * Schedules idle task to be the next runnable task on current CPU.
6155 * It does so by boosting its priority to highest possible.
6156 * Used by CPU offline code.
6158 void sched_idle_next(void)
6160 int this_cpu = smp_processor_id();
6161 struct rq *rq = cpu_rq(this_cpu);
6162 struct task_struct *p = rq->idle;
6163 unsigned long flags;
6165 /* cpu has to be offline */
6166 BUG_ON(cpu_online(this_cpu));
6169 * Strictly not necessary since rest of the CPUs are stopped by now
6170 * and interrupts disabled on the current cpu.
6172 spin_lock_irqsave(&rq->lock, flags);
6174 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6176 update_rq_clock(rq);
6177 activate_task(rq, p, 0);
6179 spin_unlock_irqrestore(&rq->lock, flags);
6183 * Ensures that the idle task is using init_mm right before its cpu goes
6186 void idle_task_exit(void)
6188 struct mm_struct *mm = current->active_mm;
6190 BUG_ON(cpu_online(smp_processor_id()));
6193 switch_mm(mm, &init_mm, current);
6197 /* called under rq->lock with disabled interrupts */
6198 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6200 struct rq *rq = cpu_rq(dead_cpu);
6202 /* Must be exiting, otherwise would be on tasklist. */
6203 BUG_ON(!p->exit_state);
6205 /* Cannot have done final schedule yet: would have vanished. */
6206 BUG_ON(p->state == TASK_DEAD);
6211 * Drop lock around migration; if someone else moves it,
6212 * that's OK. No task can be added to this CPU, so iteration is
6215 spin_unlock_irq(&rq->lock);
6216 move_task_off_dead_cpu(dead_cpu, p);
6217 spin_lock_irq(&rq->lock);
6222 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6223 static void migrate_dead_tasks(unsigned int dead_cpu)
6225 struct rq *rq = cpu_rq(dead_cpu);
6226 struct task_struct *next;
6229 if (!rq->nr_running)
6231 update_rq_clock(rq);
6232 next = pick_next_task(rq, rq->curr);
6235 migrate_dead(dead_cpu, next);
6239 #endif /* CONFIG_HOTPLUG_CPU */
6241 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6243 static struct ctl_table sd_ctl_dir[] = {
6245 .procname = "sched_domain",
6251 static struct ctl_table sd_ctl_root[] = {
6253 .ctl_name = CTL_KERN,
6254 .procname = "kernel",
6256 .child = sd_ctl_dir,
6261 static struct ctl_table *sd_alloc_ctl_entry(int n)
6263 struct ctl_table *entry =
6264 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6269 static void sd_free_ctl_entry(struct ctl_table **tablep)
6271 struct ctl_table *entry;
6274 * In the intermediate directories, both the child directory and
6275 * procname are dynamically allocated and could fail but the mode
6276 * will always be set. In the lowest directory the names are
6277 * static strings and all have proc handlers.
6279 for (entry = *tablep; entry->mode; entry++) {
6281 sd_free_ctl_entry(&entry->child);
6282 if (entry->proc_handler == NULL)
6283 kfree(entry->procname);
6291 set_table_entry(struct ctl_table *entry,
6292 const char *procname, void *data, int maxlen,
6293 mode_t mode, proc_handler *proc_handler)
6295 entry->procname = procname;
6297 entry->maxlen = maxlen;
6299 entry->proc_handler = proc_handler;
6302 static struct ctl_table *
6303 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6305 struct ctl_table *table = sd_alloc_ctl_entry(12);
6310 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6311 sizeof(long), 0644, proc_doulongvec_minmax);
6312 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6313 sizeof(long), 0644, proc_doulongvec_minmax);
6314 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6315 sizeof(int), 0644, proc_dointvec_minmax);
6316 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6317 sizeof(int), 0644, proc_dointvec_minmax);
6318 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6319 sizeof(int), 0644, proc_dointvec_minmax);
6320 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6321 sizeof(int), 0644, proc_dointvec_minmax);
6322 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6323 sizeof(int), 0644, proc_dointvec_minmax);
6324 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6325 sizeof(int), 0644, proc_dointvec_minmax);
6326 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6327 sizeof(int), 0644, proc_dointvec_minmax);
6328 set_table_entry(&table[9], "cache_nice_tries",
6329 &sd->cache_nice_tries,
6330 sizeof(int), 0644, proc_dointvec_minmax);
6331 set_table_entry(&table[10], "flags", &sd->flags,
6332 sizeof(int), 0644, proc_dointvec_minmax);
6333 /* &table[11] is terminator */
6338 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6340 struct ctl_table *entry, *table;
6341 struct sched_domain *sd;
6342 int domain_num = 0, i;
6345 for_each_domain(cpu, sd)
6347 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6352 for_each_domain(cpu, sd) {
6353 snprintf(buf, 32, "domain%d", i);
6354 entry->procname = kstrdup(buf, GFP_KERNEL);
6356 entry->child = sd_alloc_ctl_domain_table(sd);
6363 static struct ctl_table_header *sd_sysctl_header;
6364 static void register_sched_domain_sysctl(void)
6366 int i, cpu_num = num_online_cpus();
6367 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6370 WARN_ON(sd_ctl_dir[0].child);
6371 sd_ctl_dir[0].child = entry;
6376 for_each_online_cpu(i) {
6377 snprintf(buf, 32, "cpu%d", i);
6378 entry->procname = kstrdup(buf, GFP_KERNEL);
6380 entry->child = sd_alloc_ctl_cpu_table(i);
6384 WARN_ON(sd_sysctl_header);
6385 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6388 /* may be called multiple times per register */
6389 static void unregister_sched_domain_sysctl(void)
6391 if (sd_sysctl_header)
6392 unregister_sysctl_table(sd_sysctl_header);
6393 sd_sysctl_header = NULL;
6394 if (sd_ctl_dir[0].child)
6395 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6398 static void register_sched_domain_sysctl(void)
6401 static void unregister_sched_domain_sysctl(void)
6407 * migration_call - callback that gets triggered when a CPU is added.
6408 * Here we can start up the necessary migration thread for the new CPU.
6410 static int __cpuinit
6411 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6413 struct task_struct *p;
6414 int cpu = (long)hcpu;
6415 unsigned long flags;
6420 case CPU_UP_PREPARE:
6421 case CPU_UP_PREPARE_FROZEN:
6422 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6425 kthread_bind(p, cpu);
6426 /* Must be high prio: stop_machine expects to yield to it. */
6427 rq = task_rq_lock(p, &flags);
6428 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6429 task_rq_unlock(rq, &flags);
6430 cpu_rq(cpu)->migration_thread = p;
6434 case CPU_ONLINE_FROZEN:
6435 /* Strictly unnecessary, as first user will wake it. */
6436 wake_up_process(cpu_rq(cpu)->migration_thread);
6438 /* Update our root-domain */
6440 spin_lock_irqsave(&rq->lock, flags);
6442 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6443 cpu_set(cpu, rq->rd->online);
6445 spin_unlock_irqrestore(&rq->lock, flags);
6448 #ifdef CONFIG_HOTPLUG_CPU
6449 case CPU_UP_CANCELED:
6450 case CPU_UP_CANCELED_FROZEN:
6451 if (!cpu_rq(cpu)->migration_thread)
6453 /* Unbind it from offline cpu so it can run. Fall thru. */
6454 kthread_bind(cpu_rq(cpu)->migration_thread,
6455 any_online_cpu(cpu_online_map));
6456 kthread_stop(cpu_rq(cpu)->migration_thread);
6457 cpu_rq(cpu)->migration_thread = NULL;
6461 case CPU_DEAD_FROZEN:
6462 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6463 migrate_live_tasks(cpu);
6465 kthread_stop(rq->migration_thread);
6466 rq->migration_thread = NULL;
6467 /* Idle task back to normal (off runqueue, low prio) */
6468 spin_lock_irq(&rq->lock);
6469 update_rq_clock(rq);
6470 deactivate_task(rq, rq->idle, 0);
6471 rq->idle->static_prio = MAX_PRIO;
6472 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6473 rq->idle->sched_class = &idle_sched_class;
6474 migrate_dead_tasks(cpu);
6475 spin_unlock_irq(&rq->lock);
6477 migrate_nr_uninterruptible(rq);
6478 BUG_ON(rq->nr_running != 0);
6481 * No need to migrate the tasks: it was best-effort if
6482 * they didn't take sched_hotcpu_mutex. Just wake up
6485 spin_lock_irq(&rq->lock);
6486 while (!list_empty(&rq->migration_queue)) {
6487 struct migration_req *req;
6489 req = list_entry(rq->migration_queue.next,
6490 struct migration_req, list);
6491 list_del_init(&req->list);
6492 complete(&req->done);
6494 spin_unlock_irq(&rq->lock);
6498 case CPU_DYING_FROZEN:
6499 /* Update our root-domain */
6501 spin_lock_irqsave(&rq->lock, flags);
6503 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6504 cpu_clear(cpu, rq->rd->online);
6506 spin_unlock_irqrestore(&rq->lock, flags);
6513 /* Register at highest priority so that task migration (migrate_all_tasks)
6514 * happens before everything else.
6516 static struct notifier_block __cpuinitdata migration_notifier = {
6517 .notifier_call = migration_call,
6521 void __init migration_init(void)
6523 void *cpu = (void *)(long)smp_processor_id();
6526 /* Start one for the boot CPU: */
6527 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6528 BUG_ON(err == NOTIFY_BAD);
6529 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6530 register_cpu_notifier(&migration_notifier);
6536 #ifdef CONFIG_SCHED_DEBUG
6538 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6539 cpumask_t *groupmask)
6541 struct sched_group *group = sd->groups;
6544 cpulist_scnprintf(str, sizeof(str), sd->span);
6545 cpus_clear(*groupmask);
6547 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6549 if (!(sd->flags & SD_LOAD_BALANCE)) {
6550 printk("does not load-balance\n");
6552 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6557 printk(KERN_CONT "span %s\n", str);
6559 if (!cpu_isset(cpu, sd->span)) {
6560 printk(KERN_ERR "ERROR: domain->span does not contain "
6563 if (!cpu_isset(cpu, group->cpumask)) {
6564 printk(KERN_ERR "ERROR: domain->groups does not contain"
6568 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6572 printk(KERN_ERR "ERROR: group is NULL\n");
6576 if (!group->__cpu_power) {
6577 printk(KERN_CONT "\n");
6578 printk(KERN_ERR "ERROR: domain->cpu_power not "
6583 if (!cpus_weight(group->cpumask)) {
6584 printk(KERN_CONT "\n");
6585 printk(KERN_ERR "ERROR: empty group\n");
6589 if (cpus_intersects(*groupmask, group->cpumask)) {
6590 printk(KERN_CONT "\n");
6591 printk(KERN_ERR "ERROR: repeated CPUs\n");
6595 cpus_or(*groupmask, *groupmask, group->cpumask);
6597 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6598 printk(KERN_CONT " %s", str);
6600 group = group->next;
6601 } while (group != sd->groups);
6602 printk(KERN_CONT "\n");
6604 if (!cpus_equal(sd->span, *groupmask))
6605 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6607 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6608 printk(KERN_ERR "ERROR: parent span is not a superset "
6609 "of domain->span\n");
6613 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6615 cpumask_t *groupmask;
6619 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6623 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6625 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6627 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6632 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6642 # define sched_domain_debug(sd, cpu) do { } while (0)
6645 static int sd_degenerate(struct sched_domain *sd)
6647 if (cpus_weight(sd->span) == 1)
6650 /* Following flags need at least 2 groups */
6651 if (sd->flags & (SD_LOAD_BALANCE |
6652 SD_BALANCE_NEWIDLE |
6656 SD_SHARE_PKG_RESOURCES)) {
6657 if (sd->groups != sd->groups->next)
6661 /* Following flags don't use groups */
6662 if (sd->flags & (SD_WAKE_IDLE |
6671 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6673 unsigned long cflags = sd->flags, pflags = parent->flags;
6675 if (sd_degenerate(parent))
6678 if (!cpus_equal(sd->span, parent->span))
6681 /* Does parent contain flags not in child? */
6682 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6683 if (cflags & SD_WAKE_AFFINE)
6684 pflags &= ~SD_WAKE_BALANCE;
6685 /* Flags needing groups don't count if only 1 group in parent */
6686 if (parent->groups == parent->groups->next) {
6687 pflags &= ~(SD_LOAD_BALANCE |
6688 SD_BALANCE_NEWIDLE |
6692 SD_SHARE_PKG_RESOURCES);
6694 if (~cflags & pflags)
6700 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6702 unsigned long flags;
6703 const struct sched_class *class;
6705 spin_lock_irqsave(&rq->lock, flags);
6708 struct root_domain *old_rd = rq->rd;
6710 for (class = sched_class_highest; class; class = class->next) {
6711 if (class->leave_domain)
6712 class->leave_domain(rq);
6715 cpu_clear(rq->cpu, old_rd->span);
6716 cpu_clear(rq->cpu, old_rd->online);
6718 if (atomic_dec_and_test(&old_rd->refcount))
6722 atomic_inc(&rd->refcount);
6725 cpu_set(rq->cpu, rd->span);
6726 if (cpu_isset(rq->cpu, cpu_online_map))
6727 cpu_set(rq->cpu, rd->online);
6729 for (class = sched_class_highest; class; class = class->next) {
6730 if (class->join_domain)
6731 class->join_domain(rq);
6734 spin_unlock_irqrestore(&rq->lock, flags);
6737 static void init_rootdomain(struct root_domain *rd)
6739 memset(rd, 0, sizeof(*rd));
6741 cpus_clear(rd->span);
6742 cpus_clear(rd->online);
6745 static void init_defrootdomain(void)
6747 init_rootdomain(&def_root_domain);
6748 atomic_set(&def_root_domain.refcount, 1);
6751 static struct root_domain *alloc_rootdomain(void)
6753 struct root_domain *rd;
6755 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6759 init_rootdomain(rd);
6765 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6766 * hold the hotplug lock.
6769 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6771 struct rq *rq = cpu_rq(cpu);
6772 struct sched_domain *tmp;
6774 /* Remove the sched domains which do not contribute to scheduling. */
6775 for (tmp = sd; tmp; tmp = tmp->parent) {
6776 struct sched_domain *parent = tmp->parent;
6779 if (sd_parent_degenerate(tmp, parent)) {
6780 tmp->parent = parent->parent;
6782 parent->parent->child = tmp;
6786 if (sd && sd_degenerate(sd)) {
6792 sched_domain_debug(sd, cpu);
6794 rq_attach_root(rq, rd);
6795 rcu_assign_pointer(rq->sd, sd);
6798 /* cpus with isolated domains */
6799 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6801 /* Setup the mask of cpus configured for isolated domains */
6802 static int __init isolated_cpu_setup(char *str)
6804 int ints[NR_CPUS], i;
6806 str = get_options(str, ARRAY_SIZE(ints), ints);
6807 cpus_clear(cpu_isolated_map);
6808 for (i = 1; i <= ints[0]; i++)
6809 if (ints[i] < NR_CPUS)
6810 cpu_set(ints[i], cpu_isolated_map);
6814 __setup("isolcpus=", isolated_cpu_setup);
6817 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6818 * to a function which identifies what group(along with sched group) a CPU
6819 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6820 * (due to the fact that we keep track of groups covered with a cpumask_t).
6822 * init_sched_build_groups will build a circular linked list of the groups
6823 * covered by the given span, and will set each group's ->cpumask correctly,
6824 * and ->cpu_power to 0.
6827 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6828 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6829 struct sched_group **sg,
6830 cpumask_t *tmpmask),
6831 cpumask_t *covered, cpumask_t *tmpmask)
6833 struct sched_group *first = NULL, *last = NULL;
6836 cpus_clear(*covered);
6838 for_each_cpu_mask(i, *span) {
6839 struct sched_group *sg;
6840 int group = group_fn(i, cpu_map, &sg, tmpmask);
6843 if (cpu_isset(i, *covered))
6846 cpus_clear(sg->cpumask);
6847 sg->__cpu_power = 0;
6849 for_each_cpu_mask(j, *span) {
6850 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6853 cpu_set(j, *covered);
6854 cpu_set(j, sg->cpumask);
6865 #define SD_NODES_PER_DOMAIN 16
6870 * find_next_best_node - find the next node to include in a sched_domain
6871 * @node: node whose sched_domain we're building
6872 * @used_nodes: nodes already in the sched_domain
6874 * Find the next node to include in a given scheduling domain. Simply
6875 * finds the closest node not already in the @used_nodes map.
6877 * Should use nodemask_t.
6879 static int find_next_best_node(int node, nodemask_t *used_nodes)
6881 int i, n, val, min_val, best_node = 0;
6885 for (i = 0; i < MAX_NUMNODES; i++) {
6886 /* Start at @node */
6887 n = (node + i) % MAX_NUMNODES;
6889 if (!nr_cpus_node(n))
6892 /* Skip already used nodes */
6893 if (node_isset(n, *used_nodes))
6896 /* Simple min distance search */
6897 val = node_distance(node, n);
6899 if (val < min_val) {
6905 node_set(best_node, *used_nodes);
6910 * sched_domain_node_span - get a cpumask for a node's sched_domain
6911 * @node: node whose cpumask we're constructing
6912 * @span: resulting cpumask
6914 * Given a node, construct a good cpumask for its sched_domain to span. It
6915 * should be one that prevents unnecessary balancing, but also spreads tasks
6918 static void sched_domain_node_span(int node, cpumask_t *span)
6920 nodemask_t used_nodes;
6921 node_to_cpumask_ptr(nodemask, node);
6925 nodes_clear(used_nodes);
6927 cpus_or(*span, *span, *nodemask);
6928 node_set(node, used_nodes);
6930 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6931 int next_node = find_next_best_node(node, &used_nodes);
6933 node_to_cpumask_ptr_next(nodemask, next_node);
6934 cpus_or(*span, *span, *nodemask);
6939 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6942 * SMT sched-domains:
6944 #ifdef CONFIG_SCHED_SMT
6945 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6946 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6949 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6953 *sg = &per_cpu(sched_group_cpus, cpu);
6959 * multi-core sched-domains:
6961 #ifdef CONFIG_SCHED_MC
6962 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6963 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6966 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6968 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6973 *mask = per_cpu(cpu_sibling_map, cpu);
6974 cpus_and(*mask, *mask, *cpu_map);
6975 group = first_cpu(*mask);
6977 *sg = &per_cpu(sched_group_core, group);
6980 #elif defined(CONFIG_SCHED_MC)
6982 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6986 *sg = &per_cpu(sched_group_core, cpu);
6991 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6992 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6995 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6999 #ifdef CONFIG_SCHED_MC
7000 *mask = cpu_coregroup_map(cpu);
7001 cpus_and(*mask, *mask, *cpu_map);
7002 group = first_cpu(*mask);
7003 #elif defined(CONFIG_SCHED_SMT)
7004 *mask = per_cpu(cpu_sibling_map, cpu);
7005 cpus_and(*mask, *mask, *cpu_map);
7006 group = first_cpu(*mask);
7011 *sg = &per_cpu(sched_group_phys, group);
7017 * The init_sched_build_groups can't handle what we want to do with node
7018 * groups, so roll our own. Now each node has its own list of groups which
7019 * gets dynamically allocated.
7021 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7022 static struct sched_group ***sched_group_nodes_bycpu;
7024 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7025 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7027 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7028 struct sched_group **sg, cpumask_t *nodemask)
7032 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7033 cpus_and(*nodemask, *nodemask, *cpu_map);
7034 group = first_cpu(*nodemask);
7037 *sg = &per_cpu(sched_group_allnodes, group);
7041 static void init_numa_sched_groups_power(struct sched_group *group_head)
7043 struct sched_group *sg = group_head;
7049 for_each_cpu_mask(j, sg->cpumask) {
7050 struct sched_domain *sd;
7052 sd = &per_cpu(phys_domains, j);
7053 if (j != first_cpu(sd->groups->cpumask)) {
7055 * Only add "power" once for each
7061 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7064 } while (sg != group_head);
7069 /* Free memory allocated for various sched_group structures */
7070 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7074 for_each_cpu_mask(cpu, *cpu_map) {
7075 struct sched_group **sched_group_nodes
7076 = sched_group_nodes_bycpu[cpu];
7078 if (!sched_group_nodes)
7081 for (i = 0; i < MAX_NUMNODES; i++) {
7082 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7084 *nodemask = node_to_cpumask(i);
7085 cpus_and(*nodemask, *nodemask, *cpu_map);
7086 if (cpus_empty(*nodemask))
7096 if (oldsg != sched_group_nodes[i])
7099 kfree(sched_group_nodes);
7100 sched_group_nodes_bycpu[cpu] = NULL;
7104 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7110 * Initialize sched groups cpu_power.
7112 * cpu_power indicates the capacity of sched group, which is used while
7113 * distributing the load between different sched groups in a sched domain.
7114 * Typically cpu_power for all the groups in a sched domain will be same unless
7115 * there are asymmetries in the topology. If there are asymmetries, group
7116 * having more cpu_power will pickup more load compared to the group having
7119 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7120 * the maximum number of tasks a group can handle in the presence of other idle
7121 * or lightly loaded groups in the same sched domain.
7123 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7125 struct sched_domain *child;
7126 struct sched_group *group;
7128 WARN_ON(!sd || !sd->groups);
7130 if (cpu != first_cpu(sd->groups->cpumask))
7135 sd->groups->__cpu_power = 0;
7138 * For perf policy, if the groups in child domain share resources
7139 * (for example cores sharing some portions of the cache hierarchy
7140 * or SMT), then set this domain groups cpu_power such that each group
7141 * can handle only one task, when there are other idle groups in the
7142 * same sched domain.
7144 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7146 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7147 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7152 * add cpu_power of each child group to this groups cpu_power
7154 group = child->groups;
7156 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7157 group = group->next;
7158 } while (group != child->groups);
7162 * Initializers for schedule domains
7163 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7166 #define SD_INIT(sd, type) sd_init_##type(sd)
7167 #define SD_INIT_FUNC(type) \
7168 static noinline void sd_init_##type(struct sched_domain *sd) \
7170 memset(sd, 0, sizeof(*sd)); \
7171 *sd = SD_##type##_INIT; \
7172 sd->level = SD_LV_##type; \
7177 SD_INIT_FUNC(ALLNODES)
7180 #ifdef CONFIG_SCHED_SMT
7181 SD_INIT_FUNC(SIBLING)
7183 #ifdef CONFIG_SCHED_MC
7188 * To minimize stack usage kmalloc room for cpumasks and share the
7189 * space as the usage in build_sched_domains() dictates. Used only
7190 * if the amount of space is significant.
7193 cpumask_t tmpmask; /* make this one first */
7196 cpumask_t this_sibling_map;
7197 cpumask_t this_core_map;
7199 cpumask_t send_covered;
7202 cpumask_t domainspan;
7204 cpumask_t notcovered;
7209 #define SCHED_CPUMASK_ALLOC 1
7210 #define SCHED_CPUMASK_FREE(v) kfree(v)
7211 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7213 #define SCHED_CPUMASK_ALLOC 0
7214 #define SCHED_CPUMASK_FREE(v)
7215 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7218 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7219 ((unsigned long)(a) + offsetof(struct allmasks, v))
7221 static int default_relax_domain_level = -1;
7223 static int __init setup_relax_domain_level(char *str)
7225 default_relax_domain_level = simple_strtoul(str, NULL, 0);
7228 __setup("relax_domain_level=", setup_relax_domain_level);
7230 static void set_domain_attribute(struct sched_domain *sd,
7231 struct sched_domain_attr *attr)
7235 if (!attr || attr->relax_domain_level < 0) {
7236 if (default_relax_domain_level < 0)
7239 request = default_relax_domain_level;
7241 request = attr->relax_domain_level;
7242 if (request < sd->level) {
7243 /* turn off idle balance on this domain */
7244 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7246 /* turn on idle balance on this domain */
7247 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7252 * Build sched domains for a given set of cpus and attach the sched domains
7253 * to the individual cpus
7255 static int __build_sched_domains(const cpumask_t *cpu_map,
7256 struct sched_domain_attr *attr)
7259 struct root_domain *rd;
7260 SCHED_CPUMASK_DECLARE(allmasks);
7263 struct sched_group **sched_group_nodes = NULL;
7264 int sd_allnodes = 0;
7267 * Allocate the per-node list of sched groups
7269 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
7271 if (!sched_group_nodes) {
7272 printk(KERN_WARNING "Can not alloc sched group node list\n");
7277 rd = alloc_rootdomain();
7279 printk(KERN_WARNING "Cannot alloc root domain\n");
7281 kfree(sched_group_nodes);
7286 #if SCHED_CPUMASK_ALLOC
7287 /* get space for all scratch cpumask variables */
7288 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7290 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7293 kfree(sched_group_nodes);
7298 tmpmask = (cpumask_t *)allmasks;
7302 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7306 * Set up domains for cpus specified by the cpu_map.
7308 for_each_cpu_mask(i, *cpu_map) {
7309 struct sched_domain *sd = NULL, *p;
7310 SCHED_CPUMASK_VAR(nodemask, allmasks);
7312 *nodemask = node_to_cpumask(cpu_to_node(i));
7313 cpus_and(*nodemask, *nodemask, *cpu_map);
7316 if (cpus_weight(*cpu_map) >
7317 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7318 sd = &per_cpu(allnodes_domains, i);
7319 SD_INIT(sd, ALLNODES);
7320 set_domain_attribute(sd, attr);
7321 sd->span = *cpu_map;
7322 sd->first_cpu = first_cpu(sd->span);
7323 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7329 sd = &per_cpu(node_domains, i);
7331 set_domain_attribute(sd, attr);
7332 sched_domain_node_span(cpu_to_node(i), &sd->span);
7333 sd->first_cpu = first_cpu(sd->span);
7337 cpus_and(sd->span, sd->span, *cpu_map);
7341 sd = &per_cpu(phys_domains, i);
7343 set_domain_attribute(sd, attr);
7344 sd->span = *nodemask;
7345 sd->first_cpu = first_cpu(sd->span);
7349 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7351 #ifdef CONFIG_SCHED_MC
7353 sd = &per_cpu(core_domains, i);
7355 set_domain_attribute(sd, attr);
7356 sd->span = cpu_coregroup_map(i);
7357 sd->first_cpu = first_cpu(sd->span);
7358 cpus_and(sd->span, sd->span, *cpu_map);
7361 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7364 #ifdef CONFIG_SCHED_SMT
7366 sd = &per_cpu(cpu_domains, i);
7367 SD_INIT(sd, SIBLING);
7368 set_domain_attribute(sd, attr);
7369 sd->span = per_cpu(cpu_sibling_map, i);
7370 sd->first_cpu = first_cpu(sd->span);
7371 cpus_and(sd->span, sd->span, *cpu_map);
7374 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7378 #ifdef CONFIG_SCHED_SMT
7379 /* Set up CPU (sibling) groups */
7380 for_each_cpu_mask(i, *cpu_map) {
7381 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7382 SCHED_CPUMASK_VAR(send_covered, allmasks);
7384 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7385 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7386 if (i != first_cpu(*this_sibling_map))
7389 init_sched_build_groups(this_sibling_map, cpu_map,
7391 send_covered, tmpmask);
7395 #ifdef CONFIG_SCHED_MC
7396 /* Set up multi-core groups */
7397 for_each_cpu_mask(i, *cpu_map) {
7398 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7399 SCHED_CPUMASK_VAR(send_covered, allmasks);
7401 *this_core_map = cpu_coregroup_map(i);
7402 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7403 if (i != first_cpu(*this_core_map))
7406 init_sched_build_groups(this_core_map, cpu_map,
7408 send_covered, tmpmask);
7412 /* Set up physical groups */
7413 for (i = 0; i < MAX_NUMNODES; i++) {
7414 SCHED_CPUMASK_VAR(nodemask, allmasks);
7415 SCHED_CPUMASK_VAR(send_covered, allmasks);
7417 *nodemask = node_to_cpumask(i);
7418 cpus_and(*nodemask, *nodemask, *cpu_map);
7419 if (cpus_empty(*nodemask))
7422 init_sched_build_groups(nodemask, cpu_map,
7424 send_covered, tmpmask);
7428 /* Set up node groups */
7430 SCHED_CPUMASK_VAR(send_covered, allmasks);
7432 init_sched_build_groups(cpu_map, cpu_map,
7433 &cpu_to_allnodes_group,
7434 send_covered, tmpmask);
7437 for (i = 0; i < MAX_NUMNODES; i++) {
7438 /* Set up node groups */
7439 struct sched_group *sg, *prev;
7440 SCHED_CPUMASK_VAR(nodemask, allmasks);
7441 SCHED_CPUMASK_VAR(domainspan, allmasks);
7442 SCHED_CPUMASK_VAR(covered, allmasks);
7445 *nodemask = node_to_cpumask(i);
7446 cpus_clear(*covered);
7448 cpus_and(*nodemask, *nodemask, *cpu_map);
7449 if (cpus_empty(*nodemask)) {
7450 sched_group_nodes[i] = NULL;
7454 sched_domain_node_span(i, domainspan);
7455 cpus_and(*domainspan, *domainspan, *cpu_map);
7457 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7459 printk(KERN_WARNING "Can not alloc domain group for "
7463 sched_group_nodes[i] = sg;
7464 for_each_cpu_mask(j, *nodemask) {
7465 struct sched_domain *sd;
7467 sd = &per_cpu(node_domains, j);
7470 sg->__cpu_power = 0;
7471 sg->cpumask = *nodemask;
7473 cpus_or(*covered, *covered, *nodemask);
7476 for (j = 0; j < MAX_NUMNODES; j++) {
7477 SCHED_CPUMASK_VAR(notcovered, allmasks);
7478 int n = (i + j) % MAX_NUMNODES;
7479 node_to_cpumask_ptr(pnodemask, n);
7481 cpus_complement(*notcovered, *covered);
7482 cpus_and(*tmpmask, *notcovered, *cpu_map);
7483 cpus_and(*tmpmask, *tmpmask, *domainspan);
7484 if (cpus_empty(*tmpmask))
7487 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7488 if (cpus_empty(*tmpmask))
7491 sg = kmalloc_node(sizeof(struct sched_group),
7495 "Can not alloc domain group for node %d\n", j);
7498 sg->__cpu_power = 0;
7499 sg->cpumask = *tmpmask;
7500 sg->next = prev->next;
7501 cpus_or(*covered, *covered, *tmpmask);
7508 /* Calculate CPU power for physical packages and nodes */
7509 #ifdef CONFIG_SCHED_SMT
7510 for_each_cpu_mask(i, *cpu_map) {
7511 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7513 init_sched_groups_power(i, sd);
7516 #ifdef CONFIG_SCHED_MC
7517 for_each_cpu_mask(i, *cpu_map) {
7518 struct sched_domain *sd = &per_cpu(core_domains, i);
7520 init_sched_groups_power(i, sd);
7524 for_each_cpu_mask(i, *cpu_map) {
7525 struct sched_domain *sd = &per_cpu(phys_domains, i);
7527 init_sched_groups_power(i, sd);
7531 for (i = 0; i < MAX_NUMNODES; i++)
7532 init_numa_sched_groups_power(sched_group_nodes[i]);
7535 struct sched_group *sg;
7537 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7539 init_numa_sched_groups_power(sg);
7543 /* Attach the domains */
7544 for_each_cpu_mask(i, *cpu_map) {
7545 struct sched_domain *sd;
7546 #ifdef CONFIG_SCHED_SMT
7547 sd = &per_cpu(cpu_domains, i);
7548 #elif defined(CONFIG_SCHED_MC)
7549 sd = &per_cpu(core_domains, i);
7551 sd = &per_cpu(phys_domains, i);
7553 cpu_attach_domain(sd, rd, i);
7556 SCHED_CPUMASK_FREE((void *)allmasks);
7561 free_sched_groups(cpu_map, tmpmask);
7562 SCHED_CPUMASK_FREE((void *)allmasks);
7567 static int build_sched_domains(const cpumask_t *cpu_map)
7569 return __build_sched_domains(cpu_map, NULL);
7572 static cpumask_t *doms_cur; /* current sched domains */
7573 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7574 static struct sched_domain_attr *dattr_cur; /* attribues of custom domains
7578 * Special case: If a kmalloc of a doms_cur partition (array of
7579 * cpumask_t) fails, then fallback to a single sched domain,
7580 * as determined by the single cpumask_t fallback_doms.
7582 static cpumask_t fallback_doms;
7584 void __attribute__((weak)) arch_update_cpu_topology(void)
7589 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7590 * For now this just excludes isolated cpus, but could be used to
7591 * exclude other special cases in the future.
7593 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7597 arch_update_cpu_topology();
7599 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7601 doms_cur = &fallback_doms;
7602 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7604 err = build_sched_domains(doms_cur);
7605 register_sched_domain_sysctl();
7610 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7613 free_sched_groups(cpu_map, tmpmask);
7617 * Detach sched domains from a group of cpus specified in cpu_map
7618 * These cpus will now be attached to the NULL domain
7620 static void detach_destroy_domains(const cpumask_t *cpu_map)
7625 unregister_sched_domain_sysctl();
7627 for_each_cpu_mask(i, *cpu_map)
7628 cpu_attach_domain(NULL, &def_root_domain, i);
7629 synchronize_sched();
7630 arch_destroy_sched_domains(cpu_map, &tmpmask);
7633 /* handle null as "default" */
7634 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7635 struct sched_domain_attr *new, int idx_new)
7637 struct sched_domain_attr tmp;
7644 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7645 new ? (new + idx_new) : &tmp,
7646 sizeof(struct sched_domain_attr));
7650 * Partition sched domains as specified by the 'ndoms_new'
7651 * cpumasks in the array doms_new[] of cpumasks. This compares
7652 * doms_new[] to the current sched domain partitioning, doms_cur[].
7653 * It destroys each deleted domain and builds each new domain.
7655 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7656 * The masks don't intersect (don't overlap.) We should setup one
7657 * sched domain for each mask. CPUs not in any of the cpumasks will
7658 * not be load balanced. If the same cpumask appears both in the
7659 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7662 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7663 * ownership of it and will kfree it when done with it. If the caller
7664 * failed the kmalloc call, then it can pass in doms_new == NULL,
7665 * and partition_sched_domains() will fallback to the single partition
7668 * Call with hotplug lock held
7670 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7671 struct sched_domain_attr *dattr_new)
7675 mutex_lock(&sched_domains_mutex);
7677 /* always unregister in case we don't destroy any domains */
7678 unregister_sched_domain_sysctl();
7680 if (doms_new == NULL) {
7682 doms_new = &fallback_doms;
7683 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7687 /* Destroy deleted domains */
7688 for (i = 0; i < ndoms_cur; i++) {
7689 for (j = 0; j < ndoms_new; j++) {
7690 if (cpus_equal(doms_cur[i], doms_new[j])
7691 && dattrs_equal(dattr_cur, i, dattr_new, j))
7694 /* no match - a current sched domain not in new doms_new[] */
7695 detach_destroy_domains(doms_cur + i);
7700 /* Build new domains */
7701 for (i = 0; i < ndoms_new; i++) {
7702 for (j = 0; j < ndoms_cur; j++) {
7703 if (cpus_equal(doms_new[i], doms_cur[j])
7704 && dattrs_equal(dattr_new, i, dattr_cur, j))
7707 /* no match - add a new doms_new */
7708 __build_sched_domains(doms_new + i,
7709 dattr_new ? dattr_new + i : NULL);
7714 /* Remember the new sched domains */
7715 if (doms_cur != &fallback_doms)
7717 kfree(dattr_cur); /* kfree(NULL) is safe */
7718 doms_cur = doms_new;
7719 dattr_cur = dattr_new;
7720 ndoms_cur = ndoms_new;
7722 register_sched_domain_sysctl();
7724 mutex_unlock(&sched_domains_mutex);
7727 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7728 int arch_reinit_sched_domains(void)
7733 mutex_lock(&sched_domains_mutex);
7734 detach_destroy_domains(&cpu_online_map);
7735 err = arch_init_sched_domains(&cpu_online_map);
7736 mutex_unlock(&sched_domains_mutex);
7742 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7746 if (buf[0] != '0' && buf[0] != '1')
7750 sched_smt_power_savings = (buf[0] == '1');
7752 sched_mc_power_savings = (buf[0] == '1');
7754 ret = arch_reinit_sched_domains();
7756 return ret ? ret : count;
7759 #ifdef CONFIG_SCHED_MC
7760 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7762 return sprintf(page, "%u\n", sched_mc_power_savings);
7764 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7765 const char *buf, size_t count)
7767 return sched_power_savings_store(buf, count, 0);
7769 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7770 sched_mc_power_savings_store);
7773 #ifdef CONFIG_SCHED_SMT
7774 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7776 return sprintf(page, "%u\n", sched_smt_power_savings);
7778 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7779 const char *buf, size_t count)
7781 return sched_power_savings_store(buf, count, 1);
7783 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7784 sched_smt_power_savings_store);
7787 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7791 #ifdef CONFIG_SCHED_SMT
7793 err = sysfs_create_file(&cls->kset.kobj,
7794 &attr_sched_smt_power_savings.attr);
7796 #ifdef CONFIG_SCHED_MC
7797 if (!err && mc_capable())
7798 err = sysfs_create_file(&cls->kset.kobj,
7799 &attr_sched_mc_power_savings.attr);
7806 * Force a reinitialization of the sched domains hierarchy. The domains
7807 * and groups cannot be updated in place without racing with the balancing
7808 * code, so we temporarily attach all running cpus to the NULL domain
7809 * which will prevent rebalancing while the sched domains are recalculated.
7811 static int update_sched_domains(struct notifier_block *nfb,
7812 unsigned long action, void *hcpu)
7815 case CPU_UP_PREPARE:
7816 case CPU_UP_PREPARE_FROZEN:
7817 case CPU_DOWN_PREPARE:
7818 case CPU_DOWN_PREPARE_FROZEN:
7819 detach_destroy_domains(&cpu_online_map);
7822 case CPU_UP_CANCELED:
7823 case CPU_UP_CANCELED_FROZEN:
7824 case CPU_DOWN_FAILED:
7825 case CPU_DOWN_FAILED_FROZEN:
7827 case CPU_ONLINE_FROZEN:
7829 case CPU_DEAD_FROZEN:
7831 * Fall through and re-initialise the domains.
7838 /* The hotplug lock is already held by cpu_up/cpu_down */
7839 arch_init_sched_domains(&cpu_online_map);
7844 void __init sched_init_smp(void)
7846 cpumask_t non_isolated_cpus;
7848 #if defined(CONFIG_NUMA)
7849 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7851 BUG_ON(sched_group_nodes_bycpu == NULL);
7854 mutex_lock(&sched_domains_mutex);
7855 arch_init_sched_domains(&cpu_online_map);
7856 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7857 if (cpus_empty(non_isolated_cpus))
7858 cpu_set(smp_processor_id(), non_isolated_cpus);
7859 mutex_unlock(&sched_domains_mutex);
7861 /* XXX: Theoretical race here - CPU may be hotplugged now */
7862 hotcpu_notifier(update_sched_domains, 0);
7865 /* Move init over to a non-isolated CPU */
7866 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7868 sched_init_granularity();
7871 void __init sched_init_smp(void)
7873 sched_init_granularity();
7875 #endif /* CONFIG_SMP */
7877 int in_sched_functions(unsigned long addr)
7879 return in_lock_functions(addr) ||
7880 (addr >= (unsigned long)__sched_text_start
7881 && addr < (unsigned long)__sched_text_end);
7884 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7886 cfs_rq->tasks_timeline = RB_ROOT;
7887 INIT_LIST_HEAD(&cfs_rq->tasks);
7888 #ifdef CONFIG_FAIR_GROUP_SCHED
7891 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7894 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7896 struct rt_prio_array *array;
7899 array = &rt_rq->active;
7900 for (i = 0; i < MAX_RT_PRIO; i++) {
7901 INIT_LIST_HEAD(array->queue + i);
7902 __clear_bit(i, array->bitmap);
7904 /* delimiter for bitsearch: */
7905 __set_bit(MAX_RT_PRIO, array->bitmap);
7907 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7908 rt_rq->highest_prio = MAX_RT_PRIO;
7911 rt_rq->rt_nr_migratory = 0;
7912 rt_rq->overloaded = 0;
7916 rt_rq->rt_throttled = 0;
7917 rt_rq->rt_runtime = 0;
7918 spin_lock_init(&rt_rq->rt_runtime_lock);
7920 #ifdef CONFIG_RT_GROUP_SCHED
7921 rt_rq->rt_nr_boosted = 0;
7926 #ifdef CONFIG_FAIR_GROUP_SCHED
7927 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7928 struct sched_entity *se, int cpu, int add,
7929 struct sched_entity *parent)
7931 struct rq *rq = cpu_rq(cpu);
7932 tg->cfs_rq[cpu] = cfs_rq;
7933 init_cfs_rq(cfs_rq, rq);
7936 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7939 /* se could be NULL for init_task_group */
7944 se->cfs_rq = &rq->cfs;
7946 se->cfs_rq = parent->my_q;
7949 se->load.weight = tg->shares;
7950 se->load.inv_weight = 0;
7951 se->parent = parent;
7955 #ifdef CONFIG_RT_GROUP_SCHED
7956 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7957 struct sched_rt_entity *rt_se, int cpu, int add,
7958 struct sched_rt_entity *parent)
7960 struct rq *rq = cpu_rq(cpu);
7962 tg->rt_rq[cpu] = rt_rq;
7963 init_rt_rq(rt_rq, rq);
7965 rt_rq->rt_se = rt_se;
7966 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7968 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7970 tg->rt_se[cpu] = rt_se;
7975 rt_se->rt_rq = &rq->rt;
7977 rt_se->rt_rq = parent->my_q;
7979 rt_se->rt_rq = &rq->rt;
7980 rt_se->my_q = rt_rq;
7981 rt_se->parent = parent;
7982 INIT_LIST_HEAD(&rt_se->run_list);
7986 void __init sched_init(void)
7989 unsigned long alloc_size = 0, ptr;
7991 #ifdef CONFIG_FAIR_GROUP_SCHED
7992 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7994 #ifdef CONFIG_RT_GROUP_SCHED
7995 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7997 #ifdef CONFIG_USER_SCHED
8001 * As sched_init() is called before page_alloc is setup,
8002 * we use alloc_bootmem().
8005 ptr = (unsigned long)alloc_bootmem(alloc_size);
8007 #ifdef CONFIG_FAIR_GROUP_SCHED
8008 init_task_group.se = (struct sched_entity **)ptr;
8009 ptr += nr_cpu_ids * sizeof(void **);
8011 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8012 ptr += nr_cpu_ids * sizeof(void **);
8014 #ifdef CONFIG_USER_SCHED
8015 root_task_group.se = (struct sched_entity **)ptr;
8016 ptr += nr_cpu_ids * sizeof(void **);
8018 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8019 ptr += nr_cpu_ids * sizeof(void **);
8022 #ifdef CONFIG_RT_GROUP_SCHED
8023 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8024 ptr += nr_cpu_ids * sizeof(void **);
8026 init_task_group.rt_rq = (struct rt_rq **)ptr;
8027 ptr += nr_cpu_ids * sizeof(void **);
8029 #ifdef CONFIG_USER_SCHED
8030 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8031 ptr += nr_cpu_ids * sizeof(void **);
8033 root_task_group.rt_rq = (struct rt_rq **)ptr;
8034 ptr += nr_cpu_ids * sizeof(void **);
8041 init_defrootdomain();
8044 init_rt_bandwidth(&def_rt_bandwidth,
8045 global_rt_period(), global_rt_runtime());
8047 #ifdef CONFIG_RT_GROUP_SCHED
8048 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8049 global_rt_period(), global_rt_runtime());
8050 #ifdef CONFIG_USER_SCHED
8051 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8052 global_rt_period(), RUNTIME_INF);
8056 #ifdef CONFIG_GROUP_SCHED
8057 list_add(&init_task_group.list, &task_groups);
8058 INIT_LIST_HEAD(&init_task_group.children);
8060 #ifdef CONFIG_USER_SCHED
8061 INIT_LIST_HEAD(&root_task_group.children);
8062 init_task_group.parent = &root_task_group;
8063 list_add(&init_task_group.siblings, &root_task_group.children);
8067 for_each_possible_cpu(i) {
8071 spin_lock_init(&rq->lock);
8072 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
8074 init_cfs_rq(&rq->cfs, rq);
8075 init_rt_rq(&rq->rt, rq);
8076 #ifdef CONFIG_FAIR_GROUP_SCHED
8077 init_task_group.shares = init_task_group_load;
8078 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8079 #ifdef CONFIG_CGROUP_SCHED
8081 * How much cpu bandwidth does init_task_group get?
8083 * In case of task-groups formed thr' the cgroup filesystem, it
8084 * gets 100% of the cpu resources in the system. This overall
8085 * system cpu resource is divided among the tasks of
8086 * init_task_group and its child task-groups in a fair manner,
8087 * based on each entity's (task or task-group's) weight
8088 * (se->load.weight).
8090 * In other words, if init_task_group has 10 tasks of weight
8091 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8092 * then A0's share of the cpu resource is:
8094 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8096 * We achieve this by letting init_task_group's tasks sit
8097 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8099 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8100 #elif defined CONFIG_USER_SCHED
8101 root_task_group.shares = NICE_0_LOAD;
8102 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8104 * In case of task-groups formed thr' the user id of tasks,
8105 * init_task_group represents tasks belonging to root user.
8106 * Hence it forms a sibling of all subsequent groups formed.
8107 * In this case, init_task_group gets only a fraction of overall
8108 * system cpu resource, based on the weight assigned to root
8109 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8110 * by letting tasks of init_task_group sit in a separate cfs_rq
8111 * (init_cfs_rq) and having one entity represent this group of
8112 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8114 init_tg_cfs_entry(&init_task_group,
8115 &per_cpu(init_cfs_rq, i),
8116 &per_cpu(init_sched_entity, i), i, 1,
8117 root_task_group.se[i]);
8120 #endif /* CONFIG_FAIR_GROUP_SCHED */
8122 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8123 #ifdef CONFIG_RT_GROUP_SCHED
8124 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8125 #ifdef CONFIG_CGROUP_SCHED
8126 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8127 #elif defined CONFIG_USER_SCHED
8128 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8129 init_tg_rt_entry(&init_task_group,
8130 &per_cpu(init_rt_rq, i),
8131 &per_cpu(init_sched_rt_entity, i), i, 1,
8132 root_task_group.rt_se[i]);
8136 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8137 rq->cpu_load[j] = 0;
8141 rq->active_balance = 0;
8142 rq->next_balance = jiffies;
8145 rq->migration_thread = NULL;
8146 INIT_LIST_HEAD(&rq->migration_queue);
8147 rq_attach_root(rq, &def_root_domain);
8150 atomic_set(&rq->nr_iowait, 0);
8153 set_load_weight(&init_task);
8155 #ifdef CONFIG_PREEMPT_NOTIFIERS
8156 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8160 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
8163 #ifdef CONFIG_RT_MUTEXES
8164 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8168 * The boot idle thread does lazy MMU switching as well:
8170 atomic_inc(&init_mm.mm_count);
8171 enter_lazy_tlb(&init_mm, current);
8174 * Make us the idle thread. Technically, schedule() should not be
8175 * called from this thread, however somewhere below it might be,
8176 * but because we are the idle thread, we just pick up running again
8177 * when this runqueue becomes "idle".
8179 init_idle(current, smp_processor_id());
8181 * During early bootup we pretend to be a normal task:
8183 current->sched_class = &fair_sched_class;
8185 scheduler_running = 1;
8188 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8189 void __might_sleep(char *file, int line)
8192 static unsigned long prev_jiffy; /* ratelimiting */
8194 if ((in_atomic() || irqs_disabled()) &&
8195 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8196 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8198 prev_jiffy = jiffies;
8199 printk(KERN_ERR "BUG: sleeping function called from invalid"
8200 " context at %s:%d\n", file, line);
8201 printk("in_atomic():%d, irqs_disabled():%d\n",
8202 in_atomic(), irqs_disabled());
8203 debug_show_held_locks(current);
8204 if (irqs_disabled())
8205 print_irqtrace_events(current);
8210 EXPORT_SYMBOL(__might_sleep);
8213 #ifdef CONFIG_MAGIC_SYSRQ
8214 static void normalize_task(struct rq *rq, struct task_struct *p)
8218 update_rq_clock(rq);
8219 on_rq = p->se.on_rq;
8221 deactivate_task(rq, p, 0);
8222 __setscheduler(rq, p, SCHED_NORMAL, 0);
8224 activate_task(rq, p, 0);
8225 resched_task(rq->curr);
8229 void normalize_rt_tasks(void)
8231 struct task_struct *g, *p;
8232 unsigned long flags;
8235 read_lock_irqsave(&tasklist_lock, flags);
8236 do_each_thread(g, p) {
8238 * Only normalize user tasks:
8243 p->se.exec_start = 0;
8244 #ifdef CONFIG_SCHEDSTATS
8245 p->se.wait_start = 0;
8246 p->se.sleep_start = 0;
8247 p->se.block_start = 0;
8252 * Renice negative nice level userspace
8255 if (TASK_NICE(p) < 0 && p->mm)
8256 set_user_nice(p, 0);
8260 spin_lock(&p->pi_lock);
8261 rq = __task_rq_lock(p);
8263 normalize_task(rq, p);
8265 __task_rq_unlock(rq);
8266 spin_unlock(&p->pi_lock);
8267 } while_each_thread(g, p);
8269 read_unlock_irqrestore(&tasklist_lock, flags);
8272 #endif /* CONFIG_MAGIC_SYSRQ */
8276 * These functions are only useful for the IA64 MCA handling.
8278 * They can only be called when the whole system has been
8279 * stopped - every CPU needs to be quiescent, and no scheduling
8280 * activity can take place. Using them for anything else would
8281 * be a serious bug, and as a result, they aren't even visible
8282 * under any other configuration.
8286 * curr_task - return the current task for a given cpu.
8287 * @cpu: the processor in question.
8289 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8291 struct task_struct *curr_task(int cpu)
8293 return cpu_curr(cpu);
8297 * set_curr_task - set the current task for a given cpu.
8298 * @cpu: the processor in question.
8299 * @p: the task pointer to set.
8301 * Description: This function must only be used when non-maskable interrupts
8302 * are serviced on a separate stack. It allows the architecture to switch the
8303 * notion of the current task on a cpu in a non-blocking manner. This function
8304 * must be called with all CPU's synchronized, and interrupts disabled, the
8305 * and caller must save the original value of the current task (see
8306 * curr_task() above) and restore that value before reenabling interrupts and
8307 * re-starting the system.
8309 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8311 void set_curr_task(int cpu, struct task_struct *p)
8318 #ifdef CONFIG_FAIR_GROUP_SCHED
8319 static void free_fair_sched_group(struct task_group *tg)
8323 for_each_possible_cpu(i) {
8325 kfree(tg->cfs_rq[i]);
8335 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8337 struct cfs_rq *cfs_rq;
8338 struct sched_entity *se, *parent_se;
8342 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8345 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8349 tg->shares = NICE_0_LOAD;
8351 for_each_possible_cpu(i) {
8354 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8355 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8359 se = kmalloc_node(sizeof(struct sched_entity),
8360 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8364 parent_se = parent ? parent->se[i] : NULL;
8365 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8374 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8376 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8377 &cpu_rq(cpu)->leaf_cfs_rq_list);
8380 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8382 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8385 static inline void free_fair_sched_group(struct task_group *tg)
8390 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8395 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8399 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8404 #ifdef CONFIG_RT_GROUP_SCHED
8405 static void free_rt_sched_group(struct task_group *tg)
8409 destroy_rt_bandwidth(&tg->rt_bandwidth);
8411 for_each_possible_cpu(i) {
8413 kfree(tg->rt_rq[i]);
8415 kfree(tg->rt_se[i]);
8423 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8425 struct rt_rq *rt_rq;
8426 struct sched_rt_entity *rt_se, *parent_se;
8430 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8433 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8437 init_rt_bandwidth(&tg->rt_bandwidth,
8438 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8440 for_each_possible_cpu(i) {
8443 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8444 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8448 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8449 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8453 parent_se = parent ? parent->rt_se[i] : NULL;
8454 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8463 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8465 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8466 &cpu_rq(cpu)->leaf_rt_rq_list);
8469 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8471 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8474 static inline void free_rt_sched_group(struct task_group *tg)
8479 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8484 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8488 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8493 #ifdef CONFIG_GROUP_SCHED
8494 static void free_sched_group(struct task_group *tg)
8496 free_fair_sched_group(tg);
8497 free_rt_sched_group(tg);
8501 /* allocate runqueue etc for a new task group */
8502 struct task_group *sched_create_group(struct task_group *parent)
8504 struct task_group *tg;
8505 unsigned long flags;
8508 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8510 return ERR_PTR(-ENOMEM);
8512 if (!alloc_fair_sched_group(tg, parent))
8515 if (!alloc_rt_sched_group(tg, parent))
8518 spin_lock_irqsave(&task_group_lock, flags);
8519 for_each_possible_cpu(i) {
8520 register_fair_sched_group(tg, i);
8521 register_rt_sched_group(tg, i);
8523 list_add_rcu(&tg->list, &task_groups);
8525 WARN_ON(!parent); /* root should already exist */
8527 tg->parent = parent;
8528 list_add_rcu(&tg->siblings, &parent->children);
8529 INIT_LIST_HEAD(&tg->children);
8530 spin_unlock_irqrestore(&task_group_lock, flags);
8535 free_sched_group(tg);
8536 return ERR_PTR(-ENOMEM);
8539 /* rcu callback to free various structures associated with a task group */
8540 static void free_sched_group_rcu(struct rcu_head *rhp)
8542 /* now it should be safe to free those cfs_rqs */
8543 free_sched_group(container_of(rhp, struct task_group, rcu));
8546 /* Destroy runqueue etc associated with a task group */
8547 void sched_destroy_group(struct task_group *tg)
8549 unsigned long flags;
8552 spin_lock_irqsave(&task_group_lock, flags);
8553 for_each_possible_cpu(i) {
8554 unregister_fair_sched_group(tg, i);
8555 unregister_rt_sched_group(tg, i);
8557 list_del_rcu(&tg->list);
8558 list_del_rcu(&tg->siblings);
8559 spin_unlock_irqrestore(&task_group_lock, flags);
8561 /* wait for possible concurrent references to cfs_rqs complete */
8562 call_rcu(&tg->rcu, free_sched_group_rcu);
8565 /* change task's runqueue when it moves between groups.
8566 * The caller of this function should have put the task in its new group
8567 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8568 * reflect its new group.
8570 void sched_move_task(struct task_struct *tsk)
8573 unsigned long flags;
8576 rq = task_rq_lock(tsk, &flags);
8578 update_rq_clock(rq);
8580 running = task_current(rq, tsk);
8581 on_rq = tsk->se.on_rq;
8584 dequeue_task(rq, tsk, 0);
8585 if (unlikely(running))
8586 tsk->sched_class->put_prev_task(rq, tsk);
8588 set_task_rq(tsk, task_cpu(tsk));
8590 #ifdef CONFIG_FAIR_GROUP_SCHED
8591 if (tsk->sched_class->moved_group)
8592 tsk->sched_class->moved_group(tsk);
8595 if (unlikely(running))
8596 tsk->sched_class->set_curr_task(rq);
8598 enqueue_task(rq, tsk, 0);
8600 task_rq_unlock(rq, &flags);
8604 #ifdef CONFIG_FAIR_GROUP_SCHED
8605 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8607 struct cfs_rq *cfs_rq = se->cfs_rq;
8612 dequeue_entity(cfs_rq, se, 0);
8614 se->load.weight = shares;
8615 se->load.inv_weight = 0;
8618 enqueue_entity(cfs_rq, se, 0);
8621 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8623 struct cfs_rq *cfs_rq = se->cfs_rq;
8624 struct rq *rq = cfs_rq->rq;
8625 unsigned long flags;
8627 spin_lock_irqsave(&rq->lock, flags);
8628 __set_se_shares(se, shares);
8629 spin_unlock_irqrestore(&rq->lock, flags);
8632 static DEFINE_MUTEX(shares_mutex);
8634 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8637 unsigned long flags;
8640 * We can't change the weight of the root cgroup.
8645 if (shares < MIN_SHARES)
8646 shares = MIN_SHARES;
8647 else if (shares > MAX_SHARES)
8648 shares = MAX_SHARES;
8650 mutex_lock(&shares_mutex);
8651 if (tg->shares == shares)
8654 spin_lock_irqsave(&task_group_lock, flags);
8655 for_each_possible_cpu(i)
8656 unregister_fair_sched_group(tg, i);
8657 list_del_rcu(&tg->siblings);
8658 spin_unlock_irqrestore(&task_group_lock, flags);
8660 /* wait for any ongoing reference to this group to finish */
8661 synchronize_sched();
8664 * Now we are free to modify the group's share on each cpu
8665 * w/o tripping rebalance_share or load_balance_fair.
8667 tg->shares = shares;
8668 for_each_possible_cpu(i) {
8672 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8673 set_se_shares(tg->se[i], shares);
8677 * Enable load balance activity on this group, by inserting it back on
8678 * each cpu's rq->leaf_cfs_rq_list.
8680 spin_lock_irqsave(&task_group_lock, flags);
8681 for_each_possible_cpu(i)
8682 register_fair_sched_group(tg, i);
8683 list_add_rcu(&tg->siblings, &tg->parent->children);
8684 spin_unlock_irqrestore(&task_group_lock, flags);
8686 mutex_unlock(&shares_mutex);
8690 unsigned long sched_group_shares(struct task_group *tg)
8696 #ifdef CONFIG_RT_GROUP_SCHED
8698 * Ensure that the real time constraints are schedulable.
8700 static DEFINE_MUTEX(rt_constraints_mutex);
8702 static unsigned long to_ratio(u64 period, u64 runtime)
8704 if (runtime == RUNTIME_INF)
8707 return div64_u64(runtime << 16, period);
8710 #ifdef CONFIG_CGROUP_SCHED
8711 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8713 struct task_group *tgi, *parent = tg->parent;
8714 unsigned long total = 0;
8717 if (global_rt_period() < period)
8720 return to_ratio(period, runtime) <
8721 to_ratio(global_rt_period(), global_rt_runtime());
8724 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8728 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8732 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8733 tgi->rt_bandwidth.rt_runtime);
8737 return total + to_ratio(period, runtime) <
8738 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8739 parent->rt_bandwidth.rt_runtime);
8741 #elif defined CONFIG_USER_SCHED
8742 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8744 struct task_group *tgi;
8745 unsigned long total = 0;
8746 unsigned long global_ratio =
8747 to_ratio(global_rt_period(), global_rt_runtime());
8750 list_for_each_entry_rcu(tgi, &task_groups, list) {
8754 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8755 tgi->rt_bandwidth.rt_runtime);
8759 return total + to_ratio(period, runtime) < global_ratio;
8763 /* Must be called with tasklist_lock held */
8764 static inline int tg_has_rt_tasks(struct task_group *tg)
8766 struct task_struct *g, *p;
8767 do_each_thread(g, p) {
8768 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8770 } while_each_thread(g, p);
8774 static int tg_set_bandwidth(struct task_group *tg,
8775 u64 rt_period, u64 rt_runtime)
8779 mutex_lock(&rt_constraints_mutex);
8780 read_lock(&tasklist_lock);
8781 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8785 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8790 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8791 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8792 tg->rt_bandwidth.rt_runtime = rt_runtime;
8794 for_each_possible_cpu(i) {
8795 struct rt_rq *rt_rq = tg->rt_rq[i];
8797 spin_lock(&rt_rq->rt_runtime_lock);
8798 rt_rq->rt_runtime = rt_runtime;
8799 spin_unlock(&rt_rq->rt_runtime_lock);
8801 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8803 read_unlock(&tasklist_lock);
8804 mutex_unlock(&rt_constraints_mutex);
8809 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8811 u64 rt_runtime, rt_period;
8813 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8814 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8815 if (rt_runtime_us < 0)
8816 rt_runtime = RUNTIME_INF;
8818 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8821 long sched_group_rt_runtime(struct task_group *tg)
8825 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8828 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8829 do_div(rt_runtime_us, NSEC_PER_USEC);
8830 return rt_runtime_us;
8833 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8835 u64 rt_runtime, rt_period;
8837 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8838 rt_runtime = tg->rt_bandwidth.rt_runtime;
8840 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8843 long sched_group_rt_period(struct task_group *tg)
8847 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8848 do_div(rt_period_us, NSEC_PER_USEC);
8849 return rt_period_us;
8852 static int sched_rt_global_constraints(void)
8856 mutex_lock(&rt_constraints_mutex);
8857 if (!__rt_schedulable(NULL, 1, 0))
8859 mutex_unlock(&rt_constraints_mutex);
8864 static int sched_rt_global_constraints(void)
8866 unsigned long flags;
8869 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8870 for_each_possible_cpu(i) {
8871 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8873 spin_lock(&rt_rq->rt_runtime_lock);
8874 rt_rq->rt_runtime = global_rt_runtime();
8875 spin_unlock(&rt_rq->rt_runtime_lock);
8877 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8883 int sched_rt_handler(struct ctl_table *table, int write,
8884 struct file *filp, void __user *buffer, size_t *lenp,
8888 int old_period, old_runtime;
8889 static DEFINE_MUTEX(mutex);
8892 old_period = sysctl_sched_rt_period;
8893 old_runtime = sysctl_sched_rt_runtime;
8895 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8897 if (!ret && write) {
8898 ret = sched_rt_global_constraints();
8900 sysctl_sched_rt_period = old_period;
8901 sysctl_sched_rt_runtime = old_runtime;
8903 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8904 def_rt_bandwidth.rt_period =
8905 ns_to_ktime(global_rt_period());
8908 mutex_unlock(&mutex);
8913 #ifdef CONFIG_CGROUP_SCHED
8915 /* return corresponding task_group object of a cgroup */
8916 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8918 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8919 struct task_group, css);
8922 static struct cgroup_subsys_state *
8923 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8925 struct task_group *tg, *parent;
8927 if (!cgrp->parent) {
8928 /* This is early initialization for the top cgroup */
8929 init_task_group.css.cgroup = cgrp;
8930 return &init_task_group.css;
8933 parent = cgroup_tg(cgrp->parent);
8934 tg = sched_create_group(parent);
8936 return ERR_PTR(-ENOMEM);
8938 /* Bind the cgroup to task_group object we just created */
8939 tg->css.cgroup = cgrp;
8945 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8947 struct task_group *tg = cgroup_tg(cgrp);
8949 sched_destroy_group(tg);
8953 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8954 struct task_struct *tsk)
8956 #ifdef CONFIG_RT_GROUP_SCHED
8957 /* Don't accept realtime tasks when there is no way for them to run */
8958 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8961 /* We don't support RT-tasks being in separate groups */
8962 if (tsk->sched_class != &fair_sched_class)
8970 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8971 struct cgroup *old_cont, struct task_struct *tsk)
8973 sched_move_task(tsk);
8976 #ifdef CONFIG_FAIR_GROUP_SCHED
8977 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8980 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8983 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8985 struct task_group *tg = cgroup_tg(cgrp);
8987 return (u64) tg->shares;
8991 #ifdef CONFIG_RT_GROUP_SCHED
8992 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8995 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8998 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9000 return sched_group_rt_runtime(cgroup_tg(cgrp));
9003 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9006 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9009 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9011 return sched_group_rt_period(cgroup_tg(cgrp));
9015 static struct cftype cpu_files[] = {
9016 #ifdef CONFIG_FAIR_GROUP_SCHED
9019 .read_u64 = cpu_shares_read_u64,
9020 .write_u64 = cpu_shares_write_u64,
9023 #ifdef CONFIG_RT_GROUP_SCHED
9025 .name = "rt_runtime_us",
9026 .read_s64 = cpu_rt_runtime_read,
9027 .write_s64 = cpu_rt_runtime_write,
9030 .name = "rt_period_us",
9031 .read_u64 = cpu_rt_period_read_uint,
9032 .write_u64 = cpu_rt_period_write_uint,
9037 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9039 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9042 struct cgroup_subsys cpu_cgroup_subsys = {
9044 .create = cpu_cgroup_create,
9045 .destroy = cpu_cgroup_destroy,
9046 .can_attach = cpu_cgroup_can_attach,
9047 .attach = cpu_cgroup_attach,
9048 .populate = cpu_cgroup_populate,
9049 .subsys_id = cpu_cgroup_subsys_id,
9053 #endif /* CONFIG_CGROUP_SCHED */
9055 #ifdef CONFIG_CGROUP_CPUACCT
9058 * CPU accounting code for task groups.
9060 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9061 * (balbir@in.ibm.com).
9064 /* track cpu usage of a group of tasks */
9066 struct cgroup_subsys_state css;
9067 /* cpuusage holds pointer to a u64-type object on every cpu */
9071 struct cgroup_subsys cpuacct_subsys;
9073 /* return cpu accounting group corresponding to this container */
9074 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9076 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9077 struct cpuacct, css);
9080 /* return cpu accounting group to which this task belongs */
9081 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9083 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9084 struct cpuacct, css);
9087 /* create a new cpu accounting group */
9088 static struct cgroup_subsys_state *cpuacct_create(
9089 struct cgroup_subsys *ss, struct cgroup *cgrp)
9091 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9094 return ERR_PTR(-ENOMEM);
9096 ca->cpuusage = alloc_percpu(u64);
9097 if (!ca->cpuusage) {
9099 return ERR_PTR(-ENOMEM);
9105 /* destroy an existing cpu accounting group */
9107 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9109 struct cpuacct *ca = cgroup_ca(cgrp);
9111 free_percpu(ca->cpuusage);
9115 /* return total cpu usage (in nanoseconds) of a group */
9116 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9118 struct cpuacct *ca = cgroup_ca(cgrp);
9119 u64 totalcpuusage = 0;
9122 for_each_possible_cpu(i) {
9123 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9126 * Take rq->lock to make 64-bit addition safe on 32-bit
9129 spin_lock_irq(&cpu_rq(i)->lock);
9130 totalcpuusage += *cpuusage;
9131 spin_unlock_irq(&cpu_rq(i)->lock);
9134 return totalcpuusage;
9137 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9140 struct cpuacct *ca = cgroup_ca(cgrp);
9149 for_each_possible_cpu(i) {
9150 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9152 spin_lock_irq(&cpu_rq(i)->lock);
9154 spin_unlock_irq(&cpu_rq(i)->lock);
9160 static struct cftype files[] = {
9163 .read_u64 = cpuusage_read,
9164 .write_u64 = cpuusage_write,
9168 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9170 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9174 * charge this task's execution time to its accounting group.
9176 * called with rq->lock held.
9178 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9182 if (!cpuacct_subsys.active)
9187 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9189 *cpuusage += cputime;
9193 struct cgroup_subsys cpuacct_subsys = {
9195 .create = cpuacct_create,
9196 .destroy = cpuacct_destroy,
9197 .populate = cpuacct_populate,
9198 .subsys_id = cpuacct_subsys_id,
9200 #endif /* CONFIG_CGROUP_CPUACCT */