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>
77 #include "sched_cpupri.h"
80 * Convert user-nice values [ -20 ... 0 ... 19 ]
81 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
84 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
85 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
86 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
89 * 'User priority' is the nice value converted to something we
90 * can work with better when scaling various scheduler parameters,
91 * it's a [ 0 ... 39 ] range.
93 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
94 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
95 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
98 * Helpers for converting nanosecond timing to jiffy resolution
100 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
109 * Timeslices get refilled after they expire.
111 #define DEF_TIMESLICE (100 * HZ / 1000)
114 * single value that denotes runtime == period, ie unlimited time.
116 #define RUNTIME_INF ((u64)~0ULL)
120 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
121 * Since cpu_power is a 'constant', we can use a reciprocal divide.
123 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
125 return reciprocal_divide(load, sg->reciprocal_cpu_power);
129 * Each time a sched group cpu_power is changed,
130 * we must compute its reciprocal value
132 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
134 sg->__cpu_power += val;
135 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
139 static inline int rt_policy(int policy)
141 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
146 static inline int task_has_rt_policy(struct task_struct *p)
148 return rt_policy(p->policy);
152 * This is the priority-queue data structure of the RT scheduling class:
154 struct rt_prio_array {
155 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
156 struct list_head xqueue[MAX_RT_PRIO]; /* exclusive queue */
157 struct list_head squeue[MAX_RT_PRIO]; /* shared queue */
160 struct rt_bandwidth {
161 /* nests inside the rq lock: */
162 spinlock_t rt_runtime_lock;
165 struct hrtimer rt_period_timer;
168 static struct rt_bandwidth def_rt_bandwidth;
170 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
172 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
174 struct rt_bandwidth *rt_b =
175 container_of(timer, struct rt_bandwidth, rt_period_timer);
181 now = hrtimer_cb_get_time(timer);
182 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
187 idle = do_sched_rt_period_timer(rt_b, overrun);
190 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
194 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
196 rt_b->rt_period = ns_to_ktime(period);
197 rt_b->rt_runtime = runtime;
199 spin_lock_init(&rt_b->rt_runtime_lock);
201 hrtimer_init(&rt_b->rt_period_timer,
202 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
203 rt_b->rt_period_timer.function = sched_rt_period_timer;
204 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
207 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
211 if (rt_b->rt_runtime == RUNTIME_INF)
214 if (hrtimer_active(&rt_b->rt_period_timer))
217 spin_lock(&rt_b->rt_runtime_lock);
219 if (hrtimer_active(&rt_b->rt_period_timer))
222 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
223 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
224 hrtimer_start(&rt_b->rt_period_timer,
225 rt_b->rt_period_timer.expires,
228 spin_unlock(&rt_b->rt_runtime_lock);
231 #ifdef CONFIG_RT_GROUP_SCHED
232 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
234 hrtimer_cancel(&rt_b->rt_period_timer);
239 * sched_domains_mutex serializes calls to arch_init_sched_domains,
240 * detach_destroy_domains and partition_sched_domains.
242 static DEFINE_MUTEX(sched_domains_mutex);
244 #ifdef CONFIG_GROUP_SCHED
246 #include <linux/cgroup.h>
250 static LIST_HEAD(task_groups);
252 /* task group related information */
254 #ifdef CONFIG_CGROUP_SCHED
255 struct cgroup_subsys_state css;
258 #ifdef CONFIG_FAIR_GROUP_SCHED
259 /* schedulable entities of this group on each cpu */
260 struct sched_entity **se;
261 /* runqueue "owned" by this group on each cpu */
262 struct cfs_rq **cfs_rq;
263 unsigned long shares;
266 #ifdef CONFIG_RT_GROUP_SCHED
267 struct sched_rt_entity **rt_se;
268 struct rt_rq **rt_rq;
270 struct rt_bandwidth rt_bandwidth;
274 struct list_head list;
276 struct task_group *parent;
277 struct list_head siblings;
278 struct list_head children;
281 #ifdef CONFIG_USER_SCHED
285 * Every UID task group (including init_task_group aka UID-0) will
286 * be a child to this group.
288 struct task_group root_task_group;
290 #ifdef CONFIG_FAIR_GROUP_SCHED
291 /* Default task group's sched entity on each cpu */
292 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
293 /* Default task group's cfs_rq on each cpu */
294 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
297 #ifdef CONFIG_RT_GROUP_SCHED
298 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
299 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
302 #define root_task_group init_task_group
305 /* task_group_lock serializes add/remove of task groups and also changes to
306 * a task group's cpu shares.
308 static DEFINE_SPINLOCK(task_group_lock);
310 #ifdef CONFIG_FAIR_GROUP_SCHED
311 #ifdef CONFIG_USER_SCHED
312 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
314 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
318 * A weight of 0, 1 or ULONG_MAX can cause arithmetics problems.
319 * (The default weight is 1024 - so there's no practical
320 * limitation from this.)
323 #define MAX_SHARES (ULONG_MAX - 1)
325 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
328 /* Default task group.
329 * Every task in system belong to this group at bootup.
331 struct task_group init_task_group;
333 /* return group to which a task belongs */
334 static inline struct task_group *task_group(struct task_struct *p)
336 struct task_group *tg;
338 #ifdef CONFIG_USER_SCHED
340 #elif defined(CONFIG_CGROUP_SCHED)
341 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
342 struct task_group, css);
344 tg = &init_task_group;
349 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
350 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
352 #ifdef CONFIG_FAIR_GROUP_SCHED
353 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
354 p->se.parent = task_group(p)->se[cpu];
357 #ifdef CONFIG_RT_GROUP_SCHED
358 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
359 p->rt.parent = task_group(p)->rt_se[cpu];
365 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
367 #endif /* CONFIG_GROUP_SCHED */
369 /* CFS-related fields in a runqueue */
371 struct load_weight load;
372 unsigned long nr_running;
377 struct rb_root tasks_timeline;
378 struct rb_node *rb_leftmost;
380 struct list_head tasks;
381 struct list_head *balance_iterator;
384 * 'curr' points to currently running entity on this cfs_rq.
385 * It is set to NULL otherwise (i.e when none are currently running).
387 struct sched_entity *curr, *next;
389 unsigned long nr_spread_over;
391 #ifdef CONFIG_FAIR_GROUP_SCHED
392 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
395 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
396 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
397 * (like users, containers etc.)
399 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
400 * list is used during load balance.
402 struct list_head leaf_cfs_rq_list;
403 struct task_group *tg; /* group that "owns" this runqueue */
407 /* Real-Time classes' related field in a runqueue: */
409 struct rt_prio_array active;
410 unsigned long rt_nr_running;
411 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
412 int highest_prio; /* highest queued rt task prio */
415 unsigned long rt_nr_migratory;
421 /* Nests inside the rq lock: */
422 spinlock_t rt_runtime_lock;
424 #ifdef CONFIG_RT_GROUP_SCHED
425 unsigned long rt_nr_boosted;
428 struct list_head leaf_rt_rq_list;
429 struct task_group *tg;
430 struct sched_rt_entity *rt_se;
437 * We add the notion of a root-domain which will be used to define per-domain
438 * variables. Each exclusive cpuset essentially defines an island domain by
439 * fully partitioning the member cpus from any other cpuset. Whenever a new
440 * exclusive cpuset is created, we also create and attach a new root-domain
450 * The "RT overload" flag: it gets set if a CPU has more than
451 * one runnable RT task.
456 struct cpupri cpupri;
461 * By default the system creates a single root-domain with all cpus as
462 * members (mimicking the global state we have today).
464 static struct root_domain def_root_domain;
469 * This is the main, per-CPU runqueue data structure.
471 * Locking rule: those places that want to lock multiple runqueues
472 * (such as the load balancing or the thread migration code), lock
473 * acquire operations must be ordered by ascending &runqueue.
480 * nr_running and cpu_load should be in the same cacheline because
481 * remote CPUs use both these fields when doing load calculation.
483 unsigned long nr_running;
484 #define CPU_LOAD_IDX_MAX 5
485 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
486 unsigned char idle_at_tick;
488 unsigned long last_tick_seen;
489 unsigned char in_nohz_recently;
491 /* capture load from *all* tasks on this cpu: */
492 struct load_weight load;
493 unsigned long nr_load_updates;
499 #ifdef CONFIG_FAIR_GROUP_SCHED
500 /* list of leaf cfs_rq on this cpu: */
501 struct list_head leaf_cfs_rq_list;
503 #ifdef CONFIG_RT_GROUP_SCHED
504 struct list_head leaf_rt_rq_list;
508 * This is part of a global counter where only the total sum
509 * over all CPUs matters. A task can increase this counter on
510 * one CPU and if it got migrated afterwards it may decrease
511 * it on another CPU. Always updated under the runqueue lock:
513 unsigned long nr_uninterruptible;
515 struct task_struct *curr, *idle;
516 unsigned long next_balance;
517 struct mm_struct *prev_mm;
524 struct root_domain *rd;
525 struct sched_domain *sd;
527 /* For active balancing */
530 /* cpu of this runqueue: */
533 struct task_struct *migration_thread;
534 struct list_head migration_queue;
537 #ifdef CONFIG_SCHED_HRTICK
538 unsigned long hrtick_flags;
539 ktime_t hrtick_expire;
540 struct hrtimer hrtick_timer;
543 #ifdef CONFIG_SCHEDSTATS
545 struct sched_info rq_sched_info;
547 /* sys_sched_yield() stats */
548 unsigned int yld_exp_empty;
549 unsigned int yld_act_empty;
550 unsigned int yld_both_empty;
551 unsigned int yld_count;
553 /* schedule() stats */
554 unsigned int sched_switch;
555 unsigned int sched_count;
556 unsigned int sched_goidle;
558 /* try_to_wake_up() stats */
559 unsigned int ttwu_count;
560 unsigned int ttwu_local;
563 unsigned int bkl_count;
565 struct lock_class_key rq_lock_key;
568 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
570 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
572 rq->curr->sched_class->check_preempt_curr(rq, p);
575 static inline int cpu_of(struct rq *rq)
585 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
586 * See detach_destroy_domains: synchronize_sched for details.
588 * The domain tree of any CPU may only be accessed from within
589 * preempt-disabled sections.
591 #define for_each_domain(cpu, __sd) \
592 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
594 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
595 #define this_rq() (&__get_cpu_var(runqueues))
596 #define task_rq(p) cpu_rq(task_cpu(p))
597 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
599 static inline void update_rq_clock(struct rq *rq)
601 rq->clock = sched_clock_cpu(cpu_of(rq));
605 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
607 #ifdef CONFIG_SCHED_DEBUG
608 # define const_debug __read_mostly
610 # define const_debug static const
614 * Debugging: various feature bits
617 #define SCHED_FEAT(name, enabled) \
618 __SCHED_FEAT_##name ,
621 #include "sched_features.h"
626 #define SCHED_FEAT(name, enabled) \
627 (1UL << __SCHED_FEAT_##name) * enabled |
629 const_debug unsigned int sysctl_sched_features =
630 #include "sched_features.h"
635 #ifdef CONFIG_SCHED_DEBUG
636 #define SCHED_FEAT(name, enabled) \
639 static __read_mostly char *sched_feat_names[] = {
640 #include "sched_features.h"
646 static int sched_feat_open(struct inode *inode, struct file *filp)
648 filp->private_data = inode->i_private;
653 sched_feat_read(struct file *filp, char __user *ubuf,
654 size_t cnt, loff_t *ppos)
661 for (i = 0; sched_feat_names[i]; i++) {
662 len += strlen(sched_feat_names[i]);
666 buf = kmalloc(len + 2, GFP_KERNEL);
670 for (i = 0; sched_feat_names[i]; i++) {
671 if (sysctl_sched_features & (1UL << i))
672 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
674 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
677 r += sprintf(buf + r, "\n");
678 WARN_ON(r >= len + 2);
680 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
688 sched_feat_write(struct file *filp, const char __user *ubuf,
689 size_t cnt, loff_t *ppos)
699 if (copy_from_user(&buf, ubuf, cnt))
704 if (strncmp(buf, "NO_", 3) == 0) {
709 for (i = 0; sched_feat_names[i]; i++) {
710 int len = strlen(sched_feat_names[i]);
712 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
714 sysctl_sched_features &= ~(1UL << i);
716 sysctl_sched_features |= (1UL << i);
721 if (!sched_feat_names[i])
729 static struct file_operations sched_feat_fops = {
730 .open = sched_feat_open,
731 .read = sched_feat_read,
732 .write = sched_feat_write,
735 static __init int sched_init_debug(void)
737 debugfs_create_file("sched_features", 0644, NULL, NULL,
742 late_initcall(sched_init_debug);
746 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
749 * Number of tasks to iterate in a single balance run.
750 * Limited because this is done with IRQs disabled.
752 const_debug unsigned int sysctl_sched_nr_migrate = 32;
755 * period over which we measure -rt task cpu usage in us.
758 unsigned int sysctl_sched_rt_period = 1000000;
760 static __read_mostly int scheduler_running;
763 * part of the period that we allow rt tasks to run in us.
766 int sysctl_sched_rt_runtime = 950000;
768 static inline u64 global_rt_period(void)
770 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
773 static inline u64 global_rt_runtime(void)
775 if (sysctl_sched_rt_period < 0)
778 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
781 unsigned long long time_sync_thresh = 100000;
783 static DEFINE_PER_CPU(unsigned long long, time_offset);
784 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
787 * Global lock which we take every now and then to synchronize
788 * the CPUs time. This method is not warp-safe, but it's good
789 * enough to synchronize slowly diverging time sources and thus
790 * it's good enough for tracing:
792 static DEFINE_SPINLOCK(time_sync_lock);
793 static unsigned long long prev_global_time;
795 static unsigned long long __sync_cpu_clock(unsigned long long time, int cpu)
798 * We want this inlined, to not get tracer function calls
799 * in this critical section:
801 spin_acquire(&time_sync_lock.dep_map, 0, 0, _THIS_IP_);
802 __raw_spin_lock(&time_sync_lock.raw_lock);
804 if (time < prev_global_time) {
805 per_cpu(time_offset, cpu) += prev_global_time - time;
806 time = prev_global_time;
808 prev_global_time = time;
811 __raw_spin_unlock(&time_sync_lock.raw_lock);
812 spin_release(&time_sync_lock.dep_map, 1, _THIS_IP_);
817 static unsigned long long __cpu_clock(int cpu)
819 unsigned long long now;
822 * Only call sched_clock() if the scheduler has already been
823 * initialized (some code might call cpu_clock() very early):
825 if (unlikely(!scheduler_running))
828 now = sched_clock_cpu(cpu);
834 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
835 * clock constructed from sched_clock():
837 unsigned long long cpu_clock(int cpu)
839 unsigned long long prev_cpu_time, time, delta_time;
842 local_irq_save(flags);
843 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
844 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
845 delta_time = time-prev_cpu_time;
847 if (unlikely(delta_time > time_sync_thresh)) {
848 time = __sync_cpu_clock(time, cpu);
849 per_cpu(prev_cpu_time, cpu) = time;
851 local_irq_restore(flags);
855 EXPORT_SYMBOL_GPL(cpu_clock);
857 #ifndef prepare_arch_switch
858 # define prepare_arch_switch(next) do { } while (0)
860 #ifndef finish_arch_switch
861 # define finish_arch_switch(prev) do { } while (0)
864 static inline int task_current(struct rq *rq, struct task_struct *p)
866 return rq->curr == p;
869 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
870 static inline int task_running(struct rq *rq, struct task_struct *p)
872 return task_current(rq, p);
875 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
879 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
881 #ifdef CONFIG_DEBUG_SPINLOCK
882 /* this is a valid case when another task releases the spinlock */
883 rq->lock.owner = current;
886 * If we are tracking spinlock dependencies then we have to
887 * fix up the runqueue lock - which gets 'carried over' from
890 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
892 spin_unlock_irq(&rq->lock);
895 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
896 static inline int task_running(struct rq *rq, struct task_struct *p)
901 return task_current(rq, p);
905 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
909 * We can optimise this out completely for !SMP, because the
910 * SMP rebalancing from interrupt is the only thing that cares
915 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
916 spin_unlock_irq(&rq->lock);
918 spin_unlock(&rq->lock);
922 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
926 * After ->oncpu is cleared, the task can be moved to a different CPU.
927 * We must ensure this doesn't happen until the switch is completely
933 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
937 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
940 * __task_rq_lock - lock the runqueue a given task resides on.
941 * Must be called interrupts disabled.
943 static inline struct rq *__task_rq_lock(struct task_struct *p)
947 struct rq *rq = task_rq(p);
948 spin_lock(&rq->lock);
949 if (likely(rq == task_rq(p)))
951 spin_unlock(&rq->lock);
956 * task_rq_lock - lock the runqueue a given task resides on and disable
957 * interrupts. Note the ordering: we can safely lookup the task_rq without
958 * explicitly disabling preemption.
960 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
966 local_irq_save(*flags);
968 spin_lock(&rq->lock);
969 if (likely(rq == task_rq(p)))
971 spin_unlock_irqrestore(&rq->lock, *flags);
975 static void __task_rq_unlock(struct rq *rq)
978 spin_unlock(&rq->lock);
981 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
984 spin_unlock_irqrestore(&rq->lock, *flags);
988 * this_rq_lock - lock this runqueue and disable interrupts.
990 static struct rq *this_rq_lock(void)
997 spin_lock(&rq->lock);
1002 static void __resched_task(struct task_struct *p, int tif_bit);
1004 static inline void resched_task(struct task_struct *p)
1006 __resched_task(p, TIF_NEED_RESCHED);
1009 #ifdef CONFIG_SCHED_HRTICK
1011 * Use HR-timers to deliver accurate preemption points.
1013 * Its all a bit involved since we cannot program an hrt while holding the
1014 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1017 * When we get rescheduled we reprogram the hrtick_timer outside of the
1020 static inline void resched_hrt(struct task_struct *p)
1022 __resched_task(p, TIF_HRTICK_RESCHED);
1025 static inline void resched_rq(struct rq *rq)
1027 unsigned long flags;
1029 spin_lock_irqsave(&rq->lock, flags);
1030 resched_task(rq->curr);
1031 spin_unlock_irqrestore(&rq->lock, flags);
1035 HRTICK_SET, /* re-programm hrtick_timer */
1036 HRTICK_RESET, /* not a new slice */
1037 HRTICK_BLOCK, /* stop hrtick operations */
1042 * - enabled by features
1043 * - hrtimer is actually high res
1045 static inline int hrtick_enabled(struct rq *rq)
1047 if (!sched_feat(HRTICK))
1049 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
1051 return hrtimer_is_hres_active(&rq->hrtick_timer);
1055 * Called to set the hrtick timer state.
1057 * called with rq->lock held and irqs disabled
1059 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1061 assert_spin_locked(&rq->lock);
1064 * preempt at: now + delay
1067 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1069 * indicate we need to program the timer
1071 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1073 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1076 * New slices are called from the schedule path and don't need a
1077 * forced reschedule.
1080 resched_hrt(rq->curr);
1083 static void hrtick_clear(struct rq *rq)
1085 if (hrtimer_active(&rq->hrtick_timer))
1086 hrtimer_cancel(&rq->hrtick_timer);
1090 * Update the timer from the possible pending state.
1092 static void hrtick_set(struct rq *rq)
1096 unsigned long flags;
1098 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1100 spin_lock_irqsave(&rq->lock, flags);
1101 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1102 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1103 time = rq->hrtick_expire;
1104 clear_thread_flag(TIF_HRTICK_RESCHED);
1105 spin_unlock_irqrestore(&rq->lock, flags);
1108 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1109 if (reset && !hrtimer_active(&rq->hrtick_timer))
1116 * High-resolution timer tick.
1117 * Runs from hardirq context with interrupts disabled.
1119 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1121 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1123 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1125 spin_lock(&rq->lock);
1126 update_rq_clock(rq);
1127 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1128 spin_unlock(&rq->lock);
1130 return HRTIMER_NORESTART;
1133 static void hotplug_hrtick_disable(int cpu)
1135 struct rq *rq = cpu_rq(cpu);
1136 unsigned long flags;
1138 spin_lock_irqsave(&rq->lock, flags);
1139 rq->hrtick_flags = 0;
1140 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1141 spin_unlock_irqrestore(&rq->lock, flags);
1146 static void hotplug_hrtick_enable(int cpu)
1148 struct rq *rq = cpu_rq(cpu);
1149 unsigned long flags;
1151 spin_lock_irqsave(&rq->lock, flags);
1152 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1153 spin_unlock_irqrestore(&rq->lock, flags);
1157 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1159 int cpu = (int)(long)hcpu;
1162 case CPU_UP_CANCELED:
1163 case CPU_UP_CANCELED_FROZEN:
1164 case CPU_DOWN_PREPARE:
1165 case CPU_DOWN_PREPARE_FROZEN:
1167 case CPU_DEAD_FROZEN:
1168 hotplug_hrtick_disable(cpu);
1171 case CPU_UP_PREPARE:
1172 case CPU_UP_PREPARE_FROZEN:
1173 case CPU_DOWN_FAILED:
1174 case CPU_DOWN_FAILED_FROZEN:
1176 case CPU_ONLINE_FROZEN:
1177 hotplug_hrtick_enable(cpu);
1184 static void init_hrtick(void)
1186 hotcpu_notifier(hotplug_hrtick, 0);
1189 static void init_rq_hrtick(struct rq *rq)
1191 rq->hrtick_flags = 0;
1192 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1193 rq->hrtick_timer.function = hrtick;
1194 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1197 void hrtick_resched(void)
1200 unsigned long flags;
1202 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1205 local_irq_save(flags);
1206 rq = cpu_rq(smp_processor_id());
1208 local_irq_restore(flags);
1211 static inline void hrtick_clear(struct rq *rq)
1215 static inline void hrtick_set(struct rq *rq)
1219 static inline void init_rq_hrtick(struct rq *rq)
1223 void hrtick_resched(void)
1227 static inline void init_hrtick(void)
1233 * resched_task - mark a task 'to be rescheduled now'.
1235 * On UP this means the setting of the need_resched flag, on SMP it
1236 * might also involve a cross-CPU call to trigger the scheduler on
1241 #ifndef tsk_is_polling
1242 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1245 static void __resched_task(struct task_struct *p, int tif_bit)
1249 assert_spin_locked(&task_rq(p)->lock);
1251 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1254 set_tsk_thread_flag(p, tif_bit);
1257 if (cpu == smp_processor_id())
1260 /* NEED_RESCHED must be visible before we test polling */
1262 if (!tsk_is_polling(p))
1263 smp_send_reschedule(cpu);
1266 static void resched_cpu(int cpu)
1268 struct rq *rq = cpu_rq(cpu);
1269 unsigned long flags;
1271 if (!spin_trylock_irqsave(&rq->lock, flags))
1273 resched_task(cpu_curr(cpu));
1274 spin_unlock_irqrestore(&rq->lock, flags);
1279 * When add_timer_on() enqueues a timer into the timer wheel of an
1280 * idle CPU then this timer might expire before the next timer event
1281 * which is scheduled to wake up that CPU. In case of a completely
1282 * idle system the next event might even be infinite time into the
1283 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1284 * leaves the inner idle loop so the newly added timer is taken into
1285 * account when the CPU goes back to idle and evaluates the timer
1286 * wheel for the next timer event.
1288 void wake_up_idle_cpu(int cpu)
1290 struct rq *rq = cpu_rq(cpu);
1292 if (cpu == smp_processor_id())
1296 * This is safe, as this function is called with the timer
1297 * wheel base lock of (cpu) held. When the CPU is on the way
1298 * to idle and has not yet set rq->curr to idle then it will
1299 * be serialized on the timer wheel base lock and take the new
1300 * timer into account automatically.
1302 if (rq->curr != rq->idle)
1306 * We can set TIF_RESCHED on the idle task of the other CPU
1307 * lockless. The worst case is that the other CPU runs the
1308 * idle task through an additional NOOP schedule()
1310 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1312 /* NEED_RESCHED must be visible before we test polling */
1314 if (!tsk_is_polling(rq->idle))
1315 smp_send_reschedule(cpu);
1320 static void __resched_task(struct task_struct *p, int tif_bit)
1322 assert_spin_locked(&task_rq(p)->lock);
1323 set_tsk_thread_flag(p, tif_bit);
1327 #if BITS_PER_LONG == 32
1328 # define WMULT_CONST (~0UL)
1330 # define WMULT_CONST (1UL << 32)
1333 #define WMULT_SHIFT 32
1336 * Shift right and round:
1338 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1340 static unsigned long
1341 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1342 struct load_weight *lw)
1346 if (!lw->inv_weight)
1347 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)/(lw->weight+1);
1349 tmp = (u64)delta_exec * weight;
1351 * Check whether we'd overflow the 64-bit multiplication:
1353 if (unlikely(tmp > WMULT_CONST))
1354 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1357 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1359 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1362 static inline unsigned long
1363 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1365 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1368 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1374 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1381 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1382 * of tasks with abnormal "nice" values across CPUs the contribution that
1383 * each task makes to its run queue's load is weighted according to its
1384 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1385 * scaled version of the new time slice allocation that they receive on time
1389 #define WEIGHT_IDLEPRIO 2
1390 #define WMULT_IDLEPRIO (1 << 31)
1393 * Nice levels are multiplicative, with a gentle 10% change for every
1394 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1395 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1396 * that remained on nice 0.
1398 * The "10% effect" is relative and cumulative: from _any_ nice level,
1399 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1400 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1401 * If a task goes up by ~10% and another task goes down by ~10% then
1402 * the relative distance between them is ~25%.)
1404 static const int prio_to_weight[40] = {
1405 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1406 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1407 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1408 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1409 /* 0 */ 1024, 820, 655, 526, 423,
1410 /* 5 */ 335, 272, 215, 172, 137,
1411 /* 10 */ 110, 87, 70, 56, 45,
1412 /* 15 */ 36, 29, 23, 18, 15,
1416 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1418 * In cases where the weight does not change often, we can use the
1419 * precalculated inverse to speed up arithmetics by turning divisions
1420 * into multiplications:
1422 static const u32 prio_to_wmult[40] = {
1423 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1424 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1425 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1426 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1427 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1428 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1429 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1430 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1433 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1436 * runqueue iterator, to support SMP load-balancing between different
1437 * scheduling classes, without having to expose their internal data
1438 * structures to the load-balancing proper:
1440 struct rq_iterator {
1442 struct task_struct *(*start)(void *);
1443 struct task_struct *(*next)(void *);
1447 static unsigned long
1448 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1449 unsigned long max_load_move, struct sched_domain *sd,
1450 enum cpu_idle_type idle, int *all_pinned,
1451 int *this_best_prio, struct rq_iterator *iterator);
1454 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1455 struct sched_domain *sd, enum cpu_idle_type idle,
1456 struct rq_iterator *iterator);
1459 #ifdef CONFIG_CGROUP_CPUACCT
1460 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1462 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1465 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1467 update_load_add(&rq->load, load);
1470 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1472 update_load_sub(&rq->load, load);
1476 static unsigned long source_load(int cpu, int type);
1477 static unsigned long target_load(int cpu, int type);
1478 static unsigned long cpu_avg_load_per_task(int cpu);
1479 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1480 #else /* CONFIG_SMP */
1482 #ifdef CONFIG_FAIR_GROUP_SCHED
1483 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1488 #endif /* CONFIG_SMP */
1490 #include "sched_stats.h"
1491 #include "sched_idletask.c"
1492 #include "sched_fair.c"
1493 #include "sched_rt.c"
1494 #ifdef CONFIG_SCHED_DEBUG
1495 # include "sched_debug.c"
1498 #define sched_class_highest (&rt_sched_class)
1500 static inline void inc_load(struct rq *rq, const struct task_struct *p)
1502 update_load_add(&rq->load, p->se.load.weight);
1505 static inline void dec_load(struct rq *rq, const struct task_struct *p)
1507 update_load_sub(&rq->load, p->se.load.weight);
1510 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1516 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1522 static void set_load_weight(struct task_struct *p)
1524 if (task_has_rt_policy(p)) {
1525 p->se.load.weight = prio_to_weight[0] * 2;
1526 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1531 * SCHED_IDLE tasks get minimal weight:
1533 if (p->policy == SCHED_IDLE) {
1534 p->se.load.weight = WEIGHT_IDLEPRIO;
1535 p->se.load.inv_weight = WMULT_IDLEPRIO;
1539 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1540 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1543 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1545 sched_info_queued(p);
1546 p->sched_class->enqueue_task(rq, p, wakeup);
1550 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1552 p->sched_class->dequeue_task(rq, p, sleep);
1557 * __normal_prio - return the priority that is based on the static prio
1559 static inline int __normal_prio(struct task_struct *p)
1561 return p->static_prio;
1565 * Calculate the expected normal priority: i.e. priority
1566 * without taking RT-inheritance into account. Might be
1567 * boosted by interactivity modifiers. Changes upon fork,
1568 * setprio syscalls, and whenever the interactivity
1569 * estimator recalculates.
1571 static inline int normal_prio(struct task_struct *p)
1575 if (task_has_rt_policy(p))
1576 prio = MAX_RT_PRIO-1 - p->rt_priority;
1578 prio = __normal_prio(p);
1583 * Calculate the current priority, i.e. the priority
1584 * taken into account by the scheduler. This value might
1585 * be boosted by RT tasks, or might be boosted by
1586 * interactivity modifiers. Will be RT if the task got
1587 * RT-boosted. If not then it returns p->normal_prio.
1589 static int effective_prio(struct task_struct *p)
1591 p->normal_prio = normal_prio(p);
1593 * If we are RT tasks or we were boosted to RT priority,
1594 * keep the priority unchanged. Otherwise, update priority
1595 * to the normal priority:
1597 if (!rt_prio(p->prio))
1598 return p->normal_prio;
1603 * activate_task - move a task to the runqueue.
1605 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1607 if (task_contributes_to_load(p))
1608 rq->nr_uninterruptible--;
1610 enqueue_task(rq, p, wakeup);
1611 inc_nr_running(p, rq);
1615 * deactivate_task - remove a task from the runqueue.
1617 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1619 if (task_contributes_to_load(p))
1620 rq->nr_uninterruptible++;
1622 dequeue_task(rq, p, sleep);
1623 dec_nr_running(p, rq);
1627 * task_curr - is this task currently executing on a CPU?
1628 * @p: the task in question.
1630 inline int task_curr(const struct task_struct *p)
1632 return cpu_curr(task_cpu(p)) == p;
1635 /* Used instead of source_load when we know the type == 0 */
1636 unsigned long weighted_cpuload(const int cpu)
1638 return cpu_rq(cpu)->load.weight;
1641 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1643 set_task_rq(p, cpu);
1646 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1647 * successfuly executed on another CPU. We must ensure that updates of
1648 * per-task data have been completed by this moment.
1651 task_thread_info(p)->cpu = cpu;
1655 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1656 const struct sched_class *prev_class,
1657 int oldprio, int running)
1659 if (prev_class != p->sched_class) {
1660 if (prev_class->switched_from)
1661 prev_class->switched_from(rq, p, running);
1662 p->sched_class->switched_to(rq, p, running);
1664 p->sched_class->prio_changed(rq, p, oldprio, running);
1670 * Is this task likely cache-hot:
1673 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1678 * Buddy candidates are cache hot:
1680 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1683 if (p->sched_class != &fair_sched_class)
1686 if (sysctl_sched_migration_cost == -1)
1688 if (sysctl_sched_migration_cost == 0)
1691 delta = now - p->se.exec_start;
1693 return delta < (s64)sysctl_sched_migration_cost;
1697 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1699 int old_cpu = task_cpu(p);
1700 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1701 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1702 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1705 clock_offset = old_rq->clock - new_rq->clock;
1707 #ifdef CONFIG_SCHEDSTATS
1708 if (p->se.wait_start)
1709 p->se.wait_start -= clock_offset;
1710 if (p->se.sleep_start)
1711 p->se.sleep_start -= clock_offset;
1712 if (p->se.block_start)
1713 p->se.block_start -= clock_offset;
1714 if (old_cpu != new_cpu) {
1715 schedstat_inc(p, se.nr_migrations);
1716 if (task_hot(p, old_rq->clock, NULL))
1717 schedstat_inc(p, se.nr_forced2_migrations);
1720 p->se.vruntime -= old_cfsrq->min_vruntime -
1721 new_cfsrq->min_vruntime;
1723 __set_task_cpu(p, new_cpu);
1726 struct migration_req {
1727 struct list_head list;
1729 struct task_struct *task;
1732 struct completion done;
1736 * The task's runqueue lock must be held.
1737 * Returns true if you have to wait for migration thread.
1740 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1742 struct rq *rq = task_rq(p);
1745 * If the task is not on a runqueue (and not running), then
1746 * it is sufficient to simply update the task's cpu field.
1748 if (!p->se.on_rq && !task_running(rq, p)) {
1749 set_task_cpu(p, dest_cpu);
1753 init_completion(&req->done);
1755 req->dest_cpu = dest_cpu;
1756 list_add(&req->list, &rq->migration_queue);
1762 * wait_task_inactive - wait for a thread to unschedule.
1764 * The caller must ensure that the task *will* unschedule sometime soon,
1765 * else this function might spin for a *long* time. This function can't
1766 * be called with interrupts off, or it may introduce deadlock with
1767 * smp_call_function() if an IPI is sent by the same process we are
1768 * waiting to become inactive.
1770 void wait_task_inactive(struct task_struct *p)
1772 unsigned long flags;
1778 * We do the initial early heuristics without holding
1779 * any task-queue locks at all. We'll only try to get
1780 * the runqueue lock when things look like they will
1786 * If the task is actively running on another CPU
1787 * still, just relax and busy-wait without holding
1790 * NOTE! Since we don't hold any locks, it's not
1791 * even sure that "rq" stays as the right runqueue!
1792 * But we don't care, since "task_running()" will
1793 * return false if the runqueue has changed and p
1794 * is actually now running somewhere else!
1796 while (task_running(rq, p))
1800 * Ok, time to look more closely! We need the rq
1801 * lock now, to be *sure*. If we're wrong, we'll
1802 * just go back and repeat.
1804 rq = task_rq_lock(p, &flags);
1805 running = task_running(rq, p);
1806 on_rq = p->se.on_rq;
1807 task_rq_unlock(rq, &flags);
1810 * Was it really running after all now that we
1811 * checked with the proper locks actually held?
1813 * Oops. Go back and try again..
1815 if (unlikely(running)) {
1821 * It's not enough that it's not actively running,
1822 * it must be off the runqueue _entirely_, and not
1825 * So if it wa still runnable (but just not actively
1826 * running right now), it's preempted, and we should
1827 * yield - it could be a while.
1829 if (unlikely(on_rq)) {
1830 schedule_timeout_uninterruptible(1);
1835 * Ahh, all good. It wasn't running, and it wasn't
1836 * runnable, which means that it will never become
1837 * running in the future either. We're all done!
1844 * kick_process - kick a running thread to enter/exit the kernel
1845 * @p: the to-be-kicked thread
1847 * Cause a process which is running on another CPU to enter
1848 * kernel-mode, without any delay. (to get signals handled.)
1850 * NOTE: this function doesnt have to take the runqueue lock,
1851 * because all it wants to ensure is that the remote task enters
1852 * the kernel. If the IPI races and the task has been migrated
1853 * to another CPU then no harm is done and the purpose has been
1856 void kick_process(struct task_struct *p)
1862 if ((cpu != smp_processor_id()) && task_curr(p))
1863 smp_send_reschedule(cpu);
1868 * Return a low guess at the load of a migration-source cpu weighted
1869 * according to the scheduling class and "nice" value.
1871 * We want to under-estimate the load of migration sources, to
1872 * balance conservatively.
1874 static unsigned long source_load(int cpu, int type)
1876 struct rq *rq = cpu_rq(cpu);
1877 unsigned long total = weighted_cpuload(cpu);
1882 return min(rq->cpu_load[type-1], total);
1886 * Return a high guess at the load of a migration-target cpu weighted
1887 * according to the scheduling class and "nice" value.
1889 static unsigned long target_load(int cpu, int type)
1891 struct rq *rq = cpu_rq(cpu);
1892 unsigned long total = weighted_cpuload(cpu);
1897 return max(rq->cpu_load[type-1], total);
1901 * Return the average load per task on the cpu's run queue
1903 static unsigned long cpu_avg_load_per_task(int cpu)
1905 struct rq *rq = cpu_rq(cpu);
1906 unsigned long total = weighted_cpuload(cpu);
1907 unsigned long n = rq->nr_running;
1909 return n ? total / n : SCHED_LOAD_SCALE;
1913 * find_idlest_group finds and returns the least busy CPU group within the
1916 static struct sched_group *
1917 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1919 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1920 unsigned long min_load = ULONG_MAX, this_load = 0;
1921 int load_idx = sd->forkexec_idx;
1922 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1925 unsigned long load, avg_load;
1929 /* Skip over this group if it has no CPUs allowed */
1930 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1933 local_group = cpu_isset(this_cpu, group->cpumask);
1935 /* Tally up the load of all CPUs in the group */
1938 for_each_cpu_mask(i, group->cpumask) {
1939 /* Bias balancing toward cpus of our domain */
1941 load = source_load(i, load_idx);
1943 load = target_load(i, load_idx);
1948 /* Adjust by relative CPU power of the group */
1949 avg_load = sg_div_cpu_power(group,
1950 avg_load * SCHED_LOAD_SCALE);
1953 this_load = avg_load;
1955 } else if (avg_load < min_load) {
1956 min_load = avg_load;
1959 } while (group = group->next, group != sd->groups);
1961 if (!idlest || 100*this_load < imbalance*min_load)
1967 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1970 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
1973 unsigned long load, min_load = ULONG_MAX;
1977 /* Traverse only the allowed CPUs */
1978 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
1980 for_each_cpu_mask(i, *tmp) {
1981 load = weighted_cpuload(i);
1983 if (load < min_load || (load == min_load && i == this_cpu)) {
1993 * sched_balance_self: balance the current task (running on cpu) in domains
1994 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1997 * Balance, ie. select the least loaded group.
1999 * Returns the target CPU number, or the same CPU if no balancing is needed.
2001 * preempt must be disabled.
2003 static int sched_balance_self(int cpu, int flag)
2005 struct task_struct *t = current;
2006 struct sched_domain *tmp, *sd = NULL;
2008 for_each_domain(cpu, tmp) {
2010 * If power savings logic is enabled for a domain, stop there.
2012 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2014 if (tmp->flags & flag)
2019 cpumask_t span, tmpmask;
2020 struct sched_group *group;
2021 int new_cpu, weight;
2023 if (!(sd->flags & flag)) {
2029 group = find_idlest_group(sd, t, cpu);
2035 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2036 if (new_cpu == -1 || new_cpu == cpu) {
2037 /* Now try balancing at a lower domain level of cpu */
2042 /* Now try balancing at a lower domain level of new_cpu */
2045 weight = cpus_weight(span);
2046 for_each_domain(cpu, tmp) {
2047 if (weight <= cpus_weight(tmp->span))
2049 if (tmp->flags & flag)
2052 /* while loop will break here if sd == NULL */
2058 #endif /* CONFIG_SMP */
2061 * try_to_wake_up - wake up a thread
2062 * @p: the to-be-woken-up thread
2063 * @state: the mask of task states that can be woken
2064 * @sync: do a synchronous wakeup?
2066 * Put it on the run-queue if it's not already there. The "current"
2067 * thread is always on the run-queue (except when the actual
2068 * re-schedule is in progress), and as such you're allowed to do
2069 * the simpler "current->state = TASK_RUNNING" to mark yourself
2070 * runnable without the overhead of this.
2072 * returns failure only if the task is already active.
2074 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2076 int cpu, orig_cpu, this_cpu, success = 0;
2077 unsigned long flags;
2081 if (!sched_feat(SYNC_WAKEUPS))
2085 rq = task_rq_lock(p, &flags);
2086 old_state = p->state;
2087 if (!(old_state & state))
2095 this_cpu = smp_processor_id();
2098 if (unlikely(task_running(rq, p)))
2101 cpu = p->sched_class->select_task_rq(p, sync);
2102 if (cpu != orig_cpu) {
2103 set_task_cpu(p, cpu);
2104 task_rq_unlock(rq, &flags);
2105 /* might preempt at this point */
2106 rq = task_rq_lock(p, &flags);
2107 old_state = p->state;
2108 if (!(old_state & state))
2113 this_cpu = smp_processor_id();
2117 #ifdef CONFIG_SCHEDSTATS
2118 schedstat_inc(rq, ttwu_count);
2119 if (cpu == this_cpu)
2120 schedstat_inc(rq, ttwu_local);
2122 struct sched_domain *sd;
2123 for_each_domain(this_cpu, sd) {
2124 if (cpu_isset(cpu, sd->span)) {
2125 schedstat_inc(sd, ttwu_wake_remote);
2133 #endif /* CONFIG_SMP */
2134 schedstat_inc(p, se.nr_wakeups);
2136 schedstat_inc(p, se.nr_wakeups_sync);
2137 if (orig_cpu != cpu)
2138 schedstat_inc(p, se.nr_wakeups_migrate);
2139 if (cpu == this_cpu)
2140 schedstat_inc(p, se.nr_wakeups_local);
2142 schedstat_inc(p, se.nr_wakeups_remote);
2143 update_rq_clock(rq);
2144 activate_task(rq, p, 1);
2148 check_preempt_curr(rq, p);
2150 p->state = TASK_RUNNING;
2152 if (p->sched_class->task_wake_up)
2153 p->sched_class->task_wake_up(rq, p);
2156 task_rq_unlock(rq, &flags);
2161 int wake_up_process(struct task_struct *p)
2163 return try_to_wake_up(p, TASK_ALL, 0);
2165 EXPORT_SYMBOL(wake_up_process);
2167 int wake_up_state(struct task_struct *p, unsigned int state)
2169 return try_to_wake_up(p, state, 0);
2173 * Perform scheduler related setup for a newly forked process p.
2174 * p is forked by current.
2176 * __sched_fork() is basic setup used by init_idle() too:
2178 static void __sched_fork(struct task_struct *p)
2180 p->se.exec_start = 0;
2181 p->se.sum_exec_runtime = 0;
2182 p->se.prev_sum_exec_runtime = 0;
2183 p->se.last_wakeup = 0;
2184 p->se.avg_overlap = 0;
2186 #ifdef CONFIG_SCHEDSTATS
2187 p->se.wait_start = 0;
2188 p->se.sum_sleep_runtime = 0;
2189 p->se.sleep_start = 0;
2190 p->se.block_start = 0;
2191 p->se.sleep_max = 0;
2192 p->se.block_max = 0;
2194 p->se.slice_max = 0;
2198 INIT_LIST_HEAD(&p->rt.run_list);
2200 INIT_LIST_HEAD(&p->se.group_node);
2202 #ifdef CONFIG_PREEMPT_NOTIFIERS
2203 INIT_HLIST_HEAD(&p->preempt_notifiers);
2207 * We mark the process as running here, but have not actually
2208 * inserted it onto the runqueue yet. This guarantees that
2209 * nobody will actually run it, and a signal or other external
2210 * event cannot wake it up and insert it on the runqueue either.
2212 p->state = TASK_RUNNING;
2216 * fork()/clone()-time setup:
2218 void sched_fork(struct task_struct *p, int clone_flags)
2220 int cpu = get_cpu();
2225 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2227 set_task_cpu(p, cpu);
2230 * Make sure we do not leak PI boosting priority to the child:
2232 p->prio = current->normal_prio;
2233 if (!rt_prio(p->prio))
2234 p->sched_class = &fair_sched_class;
2236 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2237 if (likely(sched_info_on()))
2238 memset(&p->sched_info, 0, sizeof(p->sched_info));
2240 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2243 #ifdef CONFIG_PREEMPT
2244 /* Want to start with kernel preemption disabled. */
2245 task_thread_info(p)->preempt_count = 1;
2251 * wake_up_new_task - wake up a newly created task for the first time.
2253 * This function will do some initial scheduler statistics housekeeping
2254 * that must be done for every newly created context, then puts the task
2255 * on the runqueue and wakes it.
2257 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2259 unsigned long flags;
2262 rq = task_rq_lock(p, &flags);
2263 BUG_ON(p->state != TASK_RUNNING);
2264 update_rq_clock(rq);
2266 p->prio = effective_prio(p);
2268 if (!p->sched_class->task_new || !current->se.on_rq) {
2269 activate_task(rq, p, 0);
2272 * Let the scheduling class do new task startup
2273 * management (if any):
2275 p->sched_class->task_new(rq, p);
2276 inc_nr_running(p, rq);
2278 check_preempt_curr(rq, p);
2280 if (p->sched_class->task_wake_up)
2281 p->sched_class->task_wake_up(rq, p);
2283 task_rq_unlock(rq, &flags);
2286 #ifdef CONFIG_PREEMPT_NOTIFIERS
2289 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2290 * @notifier: notifier struct to register
2292 void preempt_notifier_register(struct preempt_notifier *notifier)
2294 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2296 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2299 * preempt_notifier_unregister - no longer interested in preemption notifications
2300 * @notifier: notifier struct to unregister
2302 * This is safe to call from within a preemption notifier.
2304 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2306 hlist_del(¬ifier->link);
2308 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2310 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2312 struct preempt_notifier *notifier;
2313 struct hlist_node *node;
2315 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2316 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2320 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2321 struct task_struct *next)
2323 struct preempt_notifier *notifier;
2324 struct hlist_node *node;
2326 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2327 notifier->ops->sched_out(notifier, next);
2332 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2337 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2338 struct task_struct *next)
2345 * prepare_task_switch - prepare to switch tasks
2346 * @rq: the runqueue preparing to switch
2347 * @prev: the current task that is being switched out
2348 * @next: the task we are going to switch to.
2350 * This is called with the rq lock held and interrupts off. It must
2351 * be paired with a subsequent finish_task_switch after the context
2354 * prepare_task_switch sets up locking and calls architecture specific
2358 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2359 struct task_struct *next)
2361 fire_sched_out_preempt_notifiers(prev, next);
2362 prepare_lock_switch(rq, next);
2363 prepare_arch_switch(next);
2367 * finish_task_switch - clean up after a task-switch
2368 * @rq: runqueue associated with task-switch
2369 * @prev: the thread we just switched away from.
2371 * finish_task_switch must be called after the context switch, paired
2372 * with a prepare_task_switch call before the context switch.
2373 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2374 * and do any other architecture-specific cleanup actions.
2376 * Note that we may have delayed dropping an mm in context_switch(). If
2377 * so, we finish that here outside of the runqueue lock. (Doing it
2378 * with the lock held can cause deadlocks; see schedule() for
2381 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2382 __releases(rq->lock)
2384 struct mm_struct *mm = rq->prev_mm;
2390 * A task struct has one reference for the use as "current".
2391 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2392 * schedule one last time. The schedule call will never return, and
2393 * the scheduled task must drop that reference.
2394 * The test for TASK_DEAD must occur while the runqueue locks are
2395 * still held, otherwise prev could be scheduled on another cpu, die
2396 * there before we look at prev->state, and then the reference would
2398 * Manfred Spraul <manfred@colorfullife.com>
2400 prev_state = prev->state;
2401 finish_arch_switch(prev);
2402 finish_lock_switch(rq, prev);
2404 if (current->sched_class->post_schedule)
2405 current->sched_class->post_schedule(rq);
2408 fire_sched_in_preempt_notifiers(current);
2411 if (unlikely(prev_state == TASK_DEAD)) {
2413 * Remove function-return probe instances associated with this
2414 * task and put them back on the free list.
2416 kprobe_flush_task(prev);
2417 put_task_struct(prev);
2422 * schedule_tail - first thing a freshly forked thread must call.
2423 * @prev: the thread we just switched away from.
2425 asmlinkage void schedule_tail(struct task_struct *prev)
2426 __releases(rq->lock)
2428 struct rq *rq = this_rq();
2430 finish_task_switch(rq, prev);
2431 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2432 /* In this case, finish_task_switch does not reenable preemption */
2435 if (current->set_child_tid)
2436 put_user(task_pid_vnr(current), current->set_child_tid);
2440 * context_switch - switch to the new MM and the new
2441 * thread's register state.
2444 context_switch(struct rq *rq, struct task_struct *prev,
2445 struct task_struct *next)
2447 struct mm_struct *mm, *oldmm;
2449 prepare_task_switch(rq, prev, next);
2451 oldmm = prev->active_mm;
2453 * For paravirt, this is coupled with an exit in switch_to to
2454 * combine the page table reload and the switch backend into
2457 arch_enter_lazy_cpu_mode();
2459 if (unlikely(!mm)) {
2460 next->active_mm = oldmm;
2461 atomic_inc(&oldmm->mm_count);
2462 enter_lazy_tlb(oldmm, next);
2464 switch_mm(oldmm, mm, next);
2466 if (unlikely(!prev->mm)) {
2467 prev->active_mm = NULL;
2468 rq->prev_mm = oldmm;
2471 * Since the runqueue lock will be released by the next
2472 * task (which is an invalid locking op but in the case
2473 * of the scheduler it's an obvious special-case), so we
2474 * do an early lockdep release here:
2476 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2477 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2480 /* Here we just switch the register state and the stack. */
2481 switch_to(prev, next, prev);
2485 * this_rq must be evaluated again because prev may have moved
2486 * CPUs since it called schedule(), thus the 'rq' on its stack
2487 * frame will be invalid.
2489 finish_task_switch(this_rq(), prev);
2493 * nr_running, nr_uninterruptible and nr_context_switches:
2495 * externally visible scheduler statistics: current number of runnable
2496 * threads, current number of uninterruptible-sleeping threads, total
2497 * number of context switches performed since bootup.
2499 unsigned long nr_running(void)
2501 unsigned long i, sum = 0;
2503 for_each_online_cpu(i)
2504 sum += cpu_rq(i)->nr_running;
2509 unsigned long nr_uninterruptible(void)
2511 unsigned long i, sum = 0;
2513 for_each_possible_cpu(i)
2514 sum += cpu_rq(i)->nr_uninterruptible;
2517 * Since we read the counters lockless, it might be slightly
2518 * inaccurate. Do not allow it to go below zero though:
2520 if (unlikely((long)sum < 0))
2526 unsigned long long nr_context_switches(void)
2529 unsigned long long sum = 0;
2531 for_each_possible_cpu(i)
2532 sum += cpu_rq(i)->nr_switches;
2537 unsigned long nr_iowait(void)
2539 unsigned long i, sum = 0;
2541 for_each_possible_cpu(i)
2542 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2547 unsigned long nr_active(void)
2549 unsigned long i, running = 0, uninterruptible = 0;
2551 for_each_online_cpu(i) {
2552 running += cpu_rq(i)->nr_running;
2553 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2556 if (unlikely((long)uninterruptible < 0))
2557 uninterruptible = 0;
2559 return running + uninterruptible;
2563 * Update rq->cpu_load[] statistics. This function is usually called every
2564 * scheduler tick (TICK_NSEC).
2566 static void update_cpu_load(struct rq *this_rq)
2568 unsigned long this_load = this_rq->load.weight;
2571 this_rq->nr_load_updates++;
2573 /* Update our load: */
2574 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2575 unsigned long old_load, new_load;
2577 /* scale is effectively 1 << i now, and >> i divides by scale */
2579 old_load = this_rq->cpu_load[i];
2580 new_load = this_load;
2582 * Round up the averaging division if load is increasing. This
2583 * prevents us from getting stuck on 9 if the load is 10, for
2586 if (new_load > old_load)
2587 new_load += scale-1;
2588 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2595 * double_rq_lock - safely lock two runqueues
2597 * Note this does not disable interrupts like task_rq_lock,
2598 * you need to do so manually before calling.
2600 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2601 __acquires(rq1->lock)
2602 __acquires(rq2->lock)
2604 BUG_ON(!irqs_disabled());
2606 spin_lock(&rq1->lock);
2607 __acquire(rq2->lock); /* Fake it out ;) */
2610 spin_lock(&rq1->lock);
2611 spin_lock(&rq2->lock);
2613 spin_lock(&rq2->lock);
2614 spin_lock(&rq1->lock);
2617 update_rq_clock(rq1);
2618 update_rq_clock(rq2);
2622 * double_rq_unlock - safely unlock two runqueues
2624 * Note this does not restore interrupts like task_rq_unlock,
2625 * you need to do so manually after calling.
2627 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2628 __releases(rq1->lock)
2629 __releases(rq2->lock)
2631 spin_unlock(&rq1->lock);
2633 spin_unlock(&rq2->lock);
2635 __release(rq2->lock);
2639 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2641 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2642 __releases(this_rq->lock)
2643 __acquires(busiest->lock)
2644 __acquires(this_rq->lock)
2648 if (unlikely(!irqs_disabled())) {
2649 /* printk() doesn't work good under rq->lock */
2650 spin_unlock(&this_rq->lock);
2653 if (unlikely(!spin_trylock(&busiest->lock))) {
2654 if (busiest < this_rq) {
2655 spin_unlock(&this_rq->lock);
2656 spin_lock(&busiest->lock);
2657 spin_lock(&this_rq->lock);
2660 spin_lock(&busiest->lock);
2666 * If dest_cpu is allowed for this process, migrate the task to it.
2667 * This is accomplished by forcing the cpu_allowed mask to only
2668 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2669 * the cpu_allowed mask is restored.
2671 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2673 struct migration_req req;
2674 unsigned long flags;
2677 rq = task_rq_lock(p, &flags);
2678 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2679 || unlikely(cpu_is_offline(dest_cpu)))
2682 /* force the process onto the specified CPU */
2683 if (migrate_task(p, dest_cpu, &req)) {
2684 /* Need to wait for migration thread (might exit: take ref). */
2685 struct task_struct *mt = rq->migration_thread;
2687 get_task_struct(mt);
2688 task_rq_unlock(rq, &flags);
2689 wake_up_process(mt);
2690 put_task_struct(mt);
2691 wait_for_completion(&req.done);
2696 task_rq_unlock(rq, &flags);
2700 * sched_exec - execve() is a valuable balancing opportunity, because at
2701 * this point the task has the smallest effective memory and cache footprint.
2703 void sched_exec(void)
2705 int new_cpu, this_cpu = get_cpu();
2706 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2708 if (new_cpu != this_cpu)
2709 sched_migrate_task(current, new_cpu);
2713 * pull_task - move a task from a remote runqueue to the local runqueue.
2714 * Both runqueues must be locked.
2716 static void pull_task(struct rq *src_rq, struct task_struct *p,
2717 struct rq *this_rq, int this_cpu)
2719 deactivate_task(src_rq, p, 0);
2720 set_task_cpu(p, this_cpu);
2721 activate_task(this_rq, p, 0);
2723 * Note that idle threads have a prio of MAX_PRIO, for this test
2724 * to be always true for them.
2726 check_preempt_curr(this_rq, p);
2730 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2733 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2734 struct sched_domain *sd, enum cpu_idle_type idle,
2738 * We do not migrate tasks that are:
2739 * 1) running (obviously), or
2740 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2741 * 3) are cache-hot on their current CPU.
2743 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2744 schedstat_inc(p, se.nr_failed_migrations_affine);
2749 if (task_running(rq, p)) {
2750 schedstat_inc(p, se.nr_failed_migrations_running);
2755 * Aggressive migration if:
2756 * 1) task is cache cold, or
2757 * 2) too many balance attempts have failed.
2760 if (!task_hot(p, rq->clock, sd) ||
2761 sd->nr_balance_failed > sd->cache_nice_tries) {
2762 #ifdef CONFIG_SCHEDSTATS
2763 if (task_hot(p, rq->clock, sd)) {
2764 schedstat_inc(sd, lb_hot_gained[idle]);
2765 schedstat_inc(p, se.nr_forced_migrations);
2771 if (task_hot(p, rq->clock, sd)) {
2772 schedstat_inc(p, se.nr_failed_migrations_hot);
2778 static unsigned long
2779 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2780 unsigned long max_load_move, struct sched_domain *sd,
2781 enum cpu_idle_type idle, int *all_pinned,
2782 int *this_best_prio, struct rq_iterator *iterator)
2784 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2785 struct task_struct *p;
2786 long rem_load_move = max_load_move;
2788 if (max_load_move == 0)
2794 * Start the load-balancing iterator:
2796 p = iterator->start(iterator->arg);
2798 if (!p || loops++ > sysctl_sched_nr_migrate)
2801 * To help distribute high priority tasks across CPUs we don't
2802 * skip a task if it will be the highest priority task (i.e. smallest
2803 * prio value) on its new queue regardless of its load weight
2805 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2806 SCHED_LOAD_SCALE_FUZZ;
2807 if ((skip_for_load && p->prio >= *this_best_prio) ||
2808 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2809 p = iterator->next(iterator->arg);
2813 pull_task(busiest, p, this_rq, this_cpu);
2815 rem_load_move -= p->se.load.weight;
2818 * We only want to steal up to the prescribed amount of weighted load.
2820 if (rem_load_move > 0) {
2821 if (p->prio < *this_best_prio)
2822 *this_best_prio = p->prio;
2823 p = iterator->next(iterator->arg);
2828 * Right now, this is one of only two places pull_task() is called,
2829 * so we can safely collect pull_task() stats here rather than
2830 * inside pull_task().
2832 schedstat_add(sd, lb_gained[idle], pulled);
2835 *all_pinned = pinned;
2837 return max_load_move - rem_load_move;
2841 * move_tasks tries to move up to max_load_move weighted load from busiest to
2842 * this_rq, as part of a balancing operation within domain "sd".
2843 * Returns 1 if successful and 0 otherwise.
2845 * Called with both runqueues locked.
2847 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2848 unsigned long max_load_move,
2849 struct sched_domain *sd, enum cpu_idle_type idle,
2852 const struct sched_class *class = sched_class_highest;
2853 unsigned long total_load_moved = 0;
2854 int this_best_prio = this_rq->curr->prio;
2858 class->load_balance(this_rq, this_cpu, busiest,
2859 max_load_move - total_load_moved,
2860 sd, idle, all_pinned, &this_best_prio);
2861 class = class->next;
2862 } while (class && max_load_move > total_load_moved);
2864 return total_load_moved > 0;
2868 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2869 struct sched_domain *sd, enum cpu_idle_type idle,
2870 struct rq_iterator *iterator)
2872 struct task_struct *p = iterator->start(iterator->arg);
2876 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2877 pull_task(busiest, p, this_rq, this_cpu);
2879 * Right now, this is only the second place pull_task()
2880 * is called, so we can safely collect pull_task()
2881 * stats here rather than inside pull_task().
2883 schedstat_inc(sd, lb_gained[idle]);
2887 p = iterator->next(iterator->arg);
2894 * move_one_task tries to move exactly one task from busiest to this_rq, as
2895 * part of active balancing operations within "domain".
2896 * Returns 1 if successful and 0 otherwise.
2898 * Called with both runqueues locked.
2900 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2901 struct sched_domain *sd, enum cpu_idle_type idle)
2903 const struct sched_class *class;
2905 for (class = sched_class_highest; class; class = class->next)
2906 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2913 * find_busiest_group finds and returns the busiest CPU group within the
2914 * domain. It calculates and returns the amount of weighted load which
2915 * should be moved to restore balance via the imbalance parameter.
2917 static struct sched_group *
2918 find_busiest_group(struct sched_domain *sd, int this_cpu,
2919 unsigned long *imbalance, enum cpu_idle_type idle,
2920 int *sd_idle, const cpumask_t *cpus, int *balance)
2922 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2923 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2924 unsigned long max_pull;
2925 unsigned long busiest_load_per_task, busiest_nr_running;
2926 unsigned long this_load_per_task, this_nr_running;
2927 int load_idx, group_imb = 0;
2928 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2929 int power_savings_balance = 1;
2930 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2931 unsigned long min_nr_running = ULONG_MAX;
2932 struct sched_group *group_min = NULL, *group_leader = NULL;
2935 max_load = this_load = total_load = total_pwr = 0;
2936 busiest_load_per_task = busiest_nr_running = 0;
2937 this_load_per_task = this_nr_running = 0;
2938 if (idle == CPU_NOT_IDLE)
2939 load_idx = sd->busy_idx;
2940 else if (idle == CPU_NEWLY_IDLE)
2941 load_idx = sd->newidle_idx;
2943 load_idx = sd->idle_idx;
2946 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2949 int __group_imb = 0;
2950 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2951 unsigned long sum_nr_running, sum_weighted_load;
2953 local_group = cpu_isset(this_cpu, group->cpumask);
2956 balance_cpu = first_cpu(group->cpumask);
2958 /* Tally up the load of all CPUs in the group */
2959 sum_weighted_load = sum_nr_running = avg_load = 0;
2961 min_cpu_load = ~0UL;
2963 for_each_cpu_mask(i, group->cpumask) {
2966 if (!cpu_isset(i, *cpus))
2971 if (*sd_idle && rq->nr_running)
2974 /* Bias balancing toward cpus of our domain */
2976 if (idle_cpu(i) && !first_idle_cpu) {
2981 load = target_load(i, load_idx);
2983 load = source_load(i, load_idx);
2984 if (load > max_cpu_load)
2985 max_cpu_load = load;
2986 if (min_cpu_load > load)
2987 min_cpu_load = load;
2991 sum_nr_running += rq->nr_running;
2992 sum_weighted_load += weighted_cpuload(i);
2996 * First idle cpu or the first cpu(busiest) in this sched group
2997 * is eligible for doing load balancing at this and above
2998 * domains. In the newly idle case, we will allow all the cpu's
2999 * to do the newly idle load balance.
3001 if (idle != CPU_NEWLY_IDLE && local_group &&
3002 balance_cpu != this_cpu && balance) {
3007 total_load += avg_load;
3008 total_pwr += group->__cpu_power;
3010 /* Adjust by relative CPU power of the group */
3011 avg_load = sg_div_cpu_power(group,
3012 avg_load * SCHED_LOAD_SCALE);
3014 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3017 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3020 this_load = avg_load;
3022 this_nr_running = sum_nr_running;
3023 this_load_per_task = sum_weighted_load;
3024 } else if (avg_load > max_load &&
3025 (sum_nr_running > group_capacity || __group_imb)) {
3026 max_load = avg_load;
3028 busiest_nr_running = sum_nr_running;
3029 busiest_load_per_task = sum_weighted_load;
3030 group_imb = __group_imb;
3033 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3035 * Busy processors will not participate in power savings
3038 if (idle == CPU_NOT_IDLE ||
3039 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3043 * If the local group is idle or completely loaded
3044 * no need to do power savings balance at this domain
3046 if (local_group && (this_nr_running >= group_capacity ||
3048 power_savings_balance = 0;
3051 * If a group is already running at full capacity or idle,
3052 * don't include that group in power savings calculations
3054 if (!power_savings_balance || sum_nr_running >= group_capacity
3059 * Calculate the group which has the least non-idle load.
3060 * This is the group from where we need to pick up the load
3063 if ((sum_nr_running < min_nr_running) ||
3064 (sum_nr_running == min_nr_running &&
3065 first_cpu(group->cpumask) <
3066 first_cpu(group_min->cpumask))) {
3068 min_nr_running = sum_nr_running;
3069 min_load_per_task = sum_weighted_load /
3074 * Calculate the group which is almost near its
3075 * capacity but still has some space to pick up some load
3076 * from other group and save more power
3078 if (sum_nr_running <= group_capacity - 1) {
3079 if (sum_nr_running > leader_nr_running ||
3080 (sum_nr_running == leader_nr_running &&
3081 first_cpu(group->cpumask) >
3082 first_cpu(group_leader->cpumask))) {
3083 group_leader = group;
3084 leader_nr_running = sum_nr_running;
3089 group = group->next;
3090 } while (group != sd->groups);
3092 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3095 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3097 if (this_load >= avg_load ||
3098 100*max_load <= sd->imbalance_pct*this_load)
3101 busiest_load_per_task /= busiest_nr_running;
3103 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3106 * We're trying to get all the cpus to the average_load, so we don't
3107 * want to push ourselves above the average load, nor do we wish to
3108 * reduce the max loaded cpu below the average load, as either of these
3109 * actions would just result in more rebalancing later, and ping-pong
3110 * tasks around. Thus we look for the minimum possible imbalance.
3111 * Negative imbalances (*we* are more loaded than anyone else) will
3112 * be counted as no imbalance for these purposes -- we can't fix that
3113 * by pulling tasks to us. Be careful of negative numbers as they'll
3114 * appear as very large values with unsigned longs.
3116 if (max_load <= busiest_load_per_task)
3120 * In the presence of smp nice balancing, certain scenarios can have
3121 * max load less than avg load(as we skip the groups at or below
3122 * its cpu_power, while calculating max_load..)
3124 if (max_load < avg_load) {
3126 goto small_imbalance;
3129 /* Don't want to pull so many tasks that a group would go idle */
3130 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3132 /* How much load to actually move to equalise the imbalance */
3133 *imbalance = min(max_pull * busiest->__cpu_power,
3134 (avg_load - this_load) * this->__cpu_power)
3138 * if *imbalance is less than the average load per runnable task
3139 * there is no gaurantee that any tasks will be moved so we'll have
3140 * a think about bumping its value to force at least one task to be
3143 if (*imbalance < busiest_load_per_task) {
3144 unsigned long tmp, pwr_now, pwr_move;
3148 pwr_move = pwr_now = 0;
3150 if (this_nr_running) {
3151 this_load_per_task /= this_nr_running;
3152 if (busiest_load_per_task > this_load_per_task)
3155 this_load_per_task = SCHED_LOAD_SCALE;
3157 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3158 busiest_load_per_task * imbn) {
3159 *imbalance = busiest_load_per_task;
3164 * OK, we don't have enough imbalance to justify moving tasks,
3165 * however we may be able to increase total CPU power used by
3169 pwr_now += busiest->__cpu_power *
3170 min(busiest_load_per_task, max_load);
3171 pwr_now += this->__cpu_power *
3172 min(this_load_per_task, this_load);
3173 pwr_now /= SCHED_LOAD_SCALE;
3175 /* Amount of load we'd subtract */
3176 tmp = sg_div_cpu_power(busiest,
3177 busiest_load_per_task * SCHED_LOAD_SCALE);
3179 pwr_move += busiest->__cpu_power *
3180 min(busiest_load_per_task, max_load - tmp);
3182 /* Amount of load we'd add */
3183 if (max_load * busiest->__cpu_power <
3184 busiest_load_per_task * SCHED_LOAD_SCALE)
3185 tmp = sg_div_cpu_power(this,
3186 max_load * busiest->__cpu_power);
3188 tmp = sg_div_cpu_power(this,
3189 busiest_load_per_task * SCHED_LOAD_SCALE);
3190 pwr_move += this->__cpu_power *
3191 min(this_load_per_task, this_load + tmp);
3192 pwr_move /= SCHED_LOAD_SCALE;
3194 /* Move if we gain throughput */
3195 if (pwr_move > pwr_now)
3196 *imbalance = busiest_load_per_task;
3202 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3203 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3206 if (this == group_leader && group_leader != group_min) {
3207 *imbalance = min_load_per_task;
3217 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3220 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3221 unsigned long imbalance, const cpumask_t *cpus)
3223 struct rq *busiest = NULL, *rq;
3224 unsigned long max_load = 0;
3227 for_each_cpu_mask(i, group->cpumask) {
3230 if (!cpu_isset(i, *cpus))
3234 wl = weighted_cpuload(i);
3236 if (rq->nr_running == 1 && wl > imbalance)
3239 if (wl > max_load) {
3249 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3250 * so long as it is large enough.
3252 #define MAX_PINNED_INTERVAL 512
3255 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3256 * tasks if there is an imbalance.
3258 static int load_balance(int this_cpu, struct rq *this_rq,
3259 struct sched_domain *sd, enum cpu_idle_type idle,
3260 int *balance, cpumask_t *cpus)
3262 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3263 struct sched_group *group;
3264 unsigned long imbalance;
3266 unsigned long flags;
3271 * When power savings policy is enabled for the parent domain, idle
3272 * sibling can pick up load irrespective of busy siblings. In this case,
3273 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3274 * portraying it as CPU_NOT_IDLE.
3276 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3277 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3280 schedstat_inc(sd, lb_count[idle]);
3283 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3290 schedstat_inc(sd, lb_nobusyg[idle]);
3294 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3296 schedstat_inc(sd, lb_nobusyq[idle]);
3300 BUG_ON(busiest == this_rq);
3302 schedstat_add(sd, lb_imbalance[idle], imbalance);
3305 if (busiest->nr_running > 1) {
3307 * Attempt to move tasks. If find_busiest_group has found
3308 * an imbalance but busiest->nr_running <= 1, the group is
3309 * still unbalanced. ld_moved simply stays zero, so it is
3310 * correctly treated as an imbalance.
3312 local_irq_save(flags);
3313 double_rq_lock(this_rq, busiest);
3314 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3315 imbalance, sd, idle, &all_pinned);
3316 double_rq_unlock(this_rq, busiest);
3317 local_irq_restore(flags);
3320 * some other cpu did the load balance for us.
3322 if (ld_moved && this_cpu != smp_processor_id())
3323 resched_cpu(this_cpu);
3325 /* All tasks on this runqueue were pinned by CPU affinity */
3326 if (unlikely(all_pinned)) {
3327 cpu_clear(cpu_of(busiest), *cpus);
3328 if (!cpus_empty(*cpus))
3335 schedstat_inc(sd, lb_failed[idle]);
3336 sd->nr_balance_failed++;
3338 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3340 spin_lock_irqsave(&busiest->lock, flags);
3342 /* don't kick the migration_thread, if the curr
3343 * task on busiest cpu can't be moved to this_cpu
3345 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3346 spin_unlock_irqrestore(&busiest->lock, flags);
3348 goto out_one_pinned;
3351 if (!busiest->active_balance) {
3352 busiest->active_balance = 1;
3353 busiest->push_cpu = this_cpu;
3356 spin_unlock_irqrestore(&busiest->lock, flags);
3358 wake_up_process(busiest->migration_thread);
3361 * We've kicked active balancing, reset the failure
3364 sd->nr_balance_failed = sd->cache_nice_tries+1;
3367 sd->nr_balance_failed = 0;
3369 if (likely(!active_balance)) {
3370 /* We were unbalanced, so reset the balancing interval */
3371 sd->balance_interval = sd->min_interval;
3374 * If we've begun active balancing, start to back off. This
3375 * case may not be covered by the all_pinned logic if there
3376 * is only 1 task on the busy runqueue (because we don't call
3379 if (sd->balance_interval < sd->max_interval)
3380 sd->balance_interval *= 2;
3383 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3384 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3389 schedstat_inc(sd, lb_balanced[idle]);
3391 sd->nr_balance_failed = 0;
3394 /* tune up the balancing interval */
3395 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3396 (sd->balance_interval < sd->max_interval))
3397 sd->balance_interval *= 2;
3399 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3400 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3406 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3407 * tasks if there is an imbalance.
3409 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3410 * this_rq is locked.
3413 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3416 struct sched_group *group;
3417 struct rq *busiest = NULL;
3418 unsigned long imbalance;
3426 * When power savings policy is enabled for the parent domain, idle
3427 * sibling can pick up load irrespective of busy siblings. In this case,
3428 * let the state of idle sibling percolate up as IDLE, instead of
3429 * portraying it as CPU_NOT_IDLE.
3431 if (sd->flags & SD_SHARE_CPUPOWER &&
3432 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3435 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3437 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3438 &sd_idle, cpus, NULL);
3440 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3444 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3446 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3450 BUG_ON(busiest == this_rq);
3452 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3455 if (busiest->nr_running > 1) {
3456 /* Attempt to move tasks */
3457 double_lock_balance(this_rq, busiest);
3458 /* this_rq->clock is already updated */
3459 update_rq_clock(busiest);
3460 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3461 imbalance, sd, CPU_NEWLY_IDLE,
3463 spin_unlock(&busiest->lock);
3465 if (unlikely(all_pinned)) {
3466 cpu_clear(cpu_of(busiest), *cpus);
3467 if (!cpus_empty(*cpus))
3473 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3474 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3475 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3478 sd->nr_balance_failed = 0;
3483 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3484 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3485 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3487 sd->nr_balance_failed = 0;
3493 * idle_balance is called by schedule() if this_cpu is about to become
3494 * idle. Attempts to pull tasks from other CPUs.
3496 static void idle_balance(int this_cpu, struct rq *this_rq)
3498 struct sched_domain *sd;
3499 int pulled_task = -1;
3500 unsigned long next_balance = jiffies + HZ;
3503 for_each_domain(this_cpu, sd) {
3504 unsigned long interval;
3506 if (!(sd->flags & SD_LOAD_BALANCE))
3509 if (sd->flags & SD_BALANCE_NEWIDLE)
3510 /* If we've pulled tasks over stop searching: */
3511 pulled_task = load_balance_newidle(this_cpu, this_rq,
3514 interval = msecs_to_jiffies(sd->balance_interval);
3515 if (time_after(next_balance, sd->last_balance + interval))
3516 next_balance = sd->last_balance + interval;
3520 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3522 * We are going idle. next_balance may be set based on
3523 * a busy processor. So reset next_balance.
3525 this_rq->next_balance = next_balance;
3530 * active_load_balance is run by migration threads. It pushes running tasks
3531 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3532 * running on each physical CPU where possible, and avoids physical /
3533 * logical imbalances.
3535 * Called with busiest_rq locked.
3537 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3539 int target_cpu = busiest_rq->push_cpu;
3540 struct sched_domain *sd;
3541 struct rq *target_rq;
3543 /* Is there any task to move? */
3544 if (busiest_rq->nr_running <= 1)
3547 target_rq = cpu_rq(target_cpu);
3550 * This condition is "impossible", if it occurs
3551 * we need to fix it. Originally reported by
3552 * Bjorn Helgaas on a 128-cpu setup.
3554 BUG_ON(busiest_rq == target_rq);
3556 /* move a task from busiest_rq to target_rq */
3557 double_lock_balance(busiest_rq, target_rq);
3558 update_rq_clock(busiest_rq);
3559 update_rq_clock(target_rq);
3561 /* Search for an sd spanning us and the target CPU. */
3562 for_each_domain(target_cpu, sd) {
3563 if ((sd->flags & SD_LOAD_BALANCE) &&
3564 cpu_isset(busiest_cpu, sd->span))
3569 schedstat_inc(sd, alb_count);
3571 if (move_one_task(target_rq, target_cpu, busiest_rq,
3573 schedstat_inc(sd, alb_pushed);
3575 schedstat_inc(sd, alb_failed);
3577 spin_unlock(&target_rq->lock);
3582 atomic_t load_balancer;
3584 } nohz ____cacheline_aligned = {
3585 .load_balancer = ATOMIC_INIT(-1),
3586 .cpu_mask = CPU_MASK_NONE,
3590 * This routine will try to nominate the ilb (idle load balancing)
3591 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3592 * load balancing on behalf of all those cpus. If all the cpus in the system
3593 * go into this tickless mode, then there will be no ilb owner (as there is
3594 * no need for one) and all the cpus will sleep till the next wakeup event
3597 * For the ilb owner, tick is not stopped. And this tick will be used
3598 * for idle load balancing. ilb owner will still be part of
3601 * While stopping the tick, this cpu will become the ilb owner if there
3602 * is no other owner. And will be the owner till that cpu becomes busy
3603 * or if all cpus in the system stop their ticks at which point
3604 * there is no need for ilb owner.
3606 * When the ilb owner becomes busy, it nominates another owner, during the
3607 * next busy scheduler_tick()
3609 int select_nohz_load_balancer(int stop_tick)
3611 int cpu = smp_processor_id();
3614 cpu_set(cpu, nohz.cpu_mask);
3615 cpu_rq(cpu)->in_nohz_recently = 1;
3618 * If we are going offline and still the leader, give up!
3620 if (cpu_is_offline(cpu) &&
3621 atomic_read(&nohz.load_balancer) == cpu) {
3622 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3627 /* time for ilb owner also to sleep */
3628 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3629 if (atomic_read(&nohz.load_balancer) == cpu)
3630 atomic_set(&nohz.load_balancer, -1);
3634 if (atomic_read(&nohz.load_balancer) == -1) {
3635 /* make me the ilb owner */
3636 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3638 } else if (atomic_read(&nohz.load_balancer) == cpu)
3641 if (!cpu_isset(cpu, nohz.cpu_mask))
3644 cpu_clear(cpu, nohz.cpu_mask);
3646 if (atomic_read(&nohz.load_balancer) == cpu)
3647 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3654 static DEFINE_SPINLOCK(balancing);
3657 * It checks each scheduling domain to see if it is due to be balanced,
3658 * and initiates a balancing operation if so.
3660 * Balancing parameters are set up in arch_init_sched_domains.
3662 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3665 struct rq *rq = cpu_rq(cpu);
3666 unsigned long interval;
3667 struct sched_domain *sd;
3668 /* Earliest time when we have to do rebalance again */
3669 unsigned long next_balance = jiffies + 60*HZ;
3670 int update_next_balance = 0;
3673 for_each_domain(cpu, sd) {
3674 if (!(sd->flags & SD_LOAD_BALANCE))
3677 interval = sd->balance_interval;
3678 if (idle != CPU_IDLE)
3679 interval *= sd->busy_factor;
3681 /* scale ms to jiffies */
3682 interval = msecs_to_jiffies(interval);
3683 if (unlikely(!interval))
3685 if (interval > HZ*NR_CPUS/10)
3686 interval = HZ*NR_CPUS/10;
3689 if (sd->flags & SD_SERIALIZE) {
3690 if (!spin_trylock(&balancing))
3694 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3695 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3697 * We've pulled tasks over so either we're no
3698 * longer idle, or one of our SMT siblings is
3701 idle = CPU_NOT_IDLE;
3703 sd->last_balance = jiffies;
3705 if (sd->flags & SD_SERIALIZE)
3706 spin_unlock(&balancing);
3708 if (time_after(next_balance, sd->last_balance + interval)) {
3709 next_balance = sd->last_balance + interval;
3710 update_next_balance = 1;
3714 * Stop the load balance at this level. There is another
3715 * CPU in our sched group which is doing load balancing more
3723 * next_balance will be updated only when there is a need.
3724 * When the cpu is attached to null domain for ex, it will not be
3727 if (likely(update_next_balance))
3728 rq->next_balance = next_balance;
3732 * run_rebalance_domains is triggered when needed from the scheduler tick.
3733 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3734 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3736 static void run_rebalance_domains(struct softirq_action *h)
3738 int this_cpu = smp_processor_id();
3739 struct rq *this_rq = cpu_rq(this_cpu);
3740 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3741 CPU_IDLE : CPU_NOT_IDLE;
3743 rebalance_domains(this_cpu, idle);
3747 * If this cpu is the owner for idle load balancing, then do the
3748 * balancing on behalf of the other idle cpus whose ticks are
3751 if (this_rq->idle_at_tick &&
3752 atomic_read(&nohz.load_balancer) == this_cpu) {
3753 cpumask_t cpus = nohz.cpu_mask;
3757 cpu_clear(this_cpu, cpus);
3758 for_each_cpu_mask(balance_cpu, cpus) {
3760 * If this cpu gets work to do, stop the load balancing
3761 * work being done for other cpus. Next load
3762 * balancing owner will pick it up.
3767 rebalance_domains(balance_cpu, CPU_IDLE);
3769 rq = cpu_rq(balance_cpu);
3770 if (time_after(this_rq->next_balance, rq->next_balance))
3771 this_rq->next_balance = rq->next_balance;
3778 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3780 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3781 * idle load balancing owner or decide to stop the periodic load balancing,
3782 * if the whole system is idle.
3784 static inline void trigger_load_balance(struct rq *rq, int cpu)
3788 * If we were in the nohz mode recently and busy at the current
3789 * scheduler tick, then check if we need to nominate new idle
3792 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3793 rq->in_nohz_recently = 0;
3795 if (atomic_read(&nohz.load_balancer) == cpu) {
3796 cpu_clear(cpu, nohz.cpu_mask);
3797 atomic_set(&nohz.load_balancer, -1);
3800 if (atomic_read(&nohz.load_balancer) == -1) {
3802 * simple selection for now: Nominate the
3803 * first cpu in the nohz list to be the next
3806 * TBD: Traverse the sched domains and nominate
3807 * the nearest cpu in the nohz.cpu_mask.
3809 int ilb = first_cpu(nohz.cpu_mask);
3811 if (ilb < nr_cpu_ids)
3817 * If this cpu is idle and doing idle load balancing for all the
3818 * cpus with ticks stopped, is it time for that to stop?
3820 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3821 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3827 * If this cpu is idle and the idle load balancing is done by
3828 * someone else, then no need raise the SCHED_SOFTIRQ
3830 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3831 cpu_isset(cpu, nohz.cpu_mask))
3834 if (time_after_eq(jiffies, rq->next_balance))
3835 raise_softirq(SCHED_SOFTIRQ);
3838 #else /* CONFIG_SMP */
3841 * on UP we do not need to balance between CPUs:
3843 static inline void idle_balance(int cpu, struct rq *rq)
3849 DEFINE_PER_CPU(struct kernel_stat, kstat);
3851 EXPORT_PER_CPU_SYMBOL(kstat);
3854 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3855 * that have not yet been banked in case the task is currently running.
3857 unsigned long long task_sched_runtime(struct task_struct *p)
3859 unsigned long flags;
3863 rq = task_rq_lock(p, &flags);
3864 ns = p->se.sum_exec_runtime;
3865 if (task_current(rq, p)) {
3866 update_rq_clock(rq);
3867 delta_exec = rq->clock - p->se.exec_start;
3868 if ((s64)delta_exec > 0)
3871 task_rq_unlock(rq, &flags);
3877 * Account user cpu time to a process.
3878 * @p: the process that the cpu time gets accounted to
3879 * @cputime: the cpu time spent in user space since the last update
3881 void account_user_time(struct task_struct *p, cputime_t cputime)
3883 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3886 p->utime = cputime_add(p->utime, cputime);
3888 /* Add user time to cpustat. */
3889 tmp = cputime_to_cputime64(cputime);
3890 if (TASK_NICE(p) > 0)
3891 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3893 cpustat->user = cputime64_add(cpustat->user, tmp);
3897 * Account guest cpu time to a process.
3898 * @p: the process that the cpu time gets accounted to
3899 * @cputime: the cpu time spent in virtual machine since the last update
3901 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3904 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3906 tmp = cputime_to_cputime64(cputime);
3908 p->utime = cputime_add(p->utime, cputime);
3909 p->gtime = cputime_add(p->gtime, cputime);
3911 cpustat->user = cputime64_add(cpustat->user, tmp);
3912 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3916 * Account scaled user cpu time to a process.
3917 * @p: the process that the cpu time gets accounted to
3918 * @cputime: the cpu time spent in user space since the last update
3920 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3922 p->utimescaled = cputime_add(p->utimescaled, cputime);
3926 * Account system cpu time to a process.
3927 * @p: the process that the cpu time gets accounted to
3928 * @hardirq_offset: the offset to subtract from hardirq_count()
3929 * @cputime: the cpu time spent in kernel space since the last update
3931 void account_system_time(struct task_struct *p, int hardirq_offset,
3934 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3935 struct rq *rq = this_rq();
3938 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3939 account_guest_time(p, cputime);
3943 p->stime = cputime_add(p->stime, cputime);
3945 /* Add system time to cpustat. */
3946 tmp = cputime_to_cputime64(cputime);
3947 if (hardirq_count() - hardirq_offset)
3948 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3949 else if (softirq_count())
3950 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3951 else if (p != rq->idle)
3952 cpustat->system = cputime64_add(cpustat->system, tmp);
3953 else if (atomic_read(&rq->nr_iowait) > 0)
3954 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3956 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3957 /* Account for system time used */
3958 acct_update_integrals(p);
3962 * Account scaled system cpu time to a process.
3963 * @p: the process that the cpu time gets accounted to
3964 * @hardirq_offset: the offset to subtract from hardirq_count()
3965 * @cputime: the cpu time spent in kernel space since the last update
3967 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3969 p->stimescaled = cputime_add(p->stimescaled, cputime);
3973 * Account for involuntary wait time.
3974 * @p: the process from which the cpu time has been stolen
3975 * @steal: the cpu time spent in involuntary wait
3977 void account_steal_time(struct task_struct *p, cputime_t steal)
3979 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3980 cputime64_t tmp = cputime_to_cputime64(steal);
3981 struct rq *rq = this_rq();
3983 if (p == rq->idle) {
3984 p->stime = cputime_add(p->stime, steal);
3985 if (atomic_read(&rq->nr_iowait) > 0)
3986 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3988 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3990 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3994 * This function gets called by the timer code, with HZ frequency.
3995 * We call it with interrupts disabled.
3997 * It also gets called by the fork code, when changing the parent's
4000 void scheduler_tick(void)
4002 int cpu = smp_processor_id();
4003 struct rq *rq = cpu_rq(cpu);
4004 struct task_struct *curr = rq->curr;
4008 spin_lock(&rq->lock);
4009 update_rq_clock(rq);
4010 update_cpu_load(rq);
4011 curr->sched_class->task_tick(rq, curr, 0);
4012 spin_unlock(&rq->lock);
4015 rq->idle_at_tick = idle_cpu(cpu);
4016 trigger_load_balance(rq, cpu);
4020 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4022 void __kprobes add_preempt_count(int val)
4027 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4029 preempt_count() += val;
4031 * Spinlock count overflowing soon?
4033 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4036 EXPORT_SYMBOL(add_preempt_count);
4038 void __kprobes sub_preempt_count(int val)
4043 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4046 * Is the spinlock portion underflowing?
4048 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4049 !(preempt_count() & PREEMPT_MASK)))
4052 preempt_count() -= val;
4054 EXPORT_SYMBOL(sub_preempt_count);
4059 * Print scheduling while atomic bug:
4061 static noinline void __schedule_bug(struct task_struct *prev)
4063 struct pt_regs *regs = get_irq_regs();
4065 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4066 prev->comm, prev->pid, preempt_count());
4068 debug_show_held_locks(prev);
4069 if (irqs_disabled())
4070 print_irqtrace_events(prev);
4079 * Various schedule()-time debugging checks and statistics:
4081 static inline void schedule_debug(struct task_struct *prev)
4084 * Test if we are atomic. Since do_exit() needs to call into
4085 * schedule() atomically, we ignore that path for now.
4086 * Otherwise, whine if we are scheduling when we should not be.
4088 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4089 __schedule_bug(prev);
4091 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4093 schedstat_inc(this_rq(), sched_count);
4094 #ifdef CONFIG_SCHEDSTATS
4095 if (unlikely(prev->lock_depth >= 0)) {
4096 schedstat_inc(this_rq(), bkl_count);
4097 schedstat_inc(prev, sched_info.bkl_count);
4103 * Pick up the highest-prio task:
4105 static inline struct task_struct *
4106 pick_next_task(struct rq *rq, struct task_struct *prev)
4108 const struct sched_class *class;
4109 struct task_struct *p;
4112 * Optimization: we know that if all tasks are in
4113 * the fair class we can call that function directly:
4115 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4116 p = fair_sched_class.pick_next_task(rq);
4121 class = sched_class_highest;
4123 p = class->pick_next_task(rq);
4127 * Will never be NULL as the idle class always
4128 * returns a non-NULL p:
4130 class = class->next;
4135 * schedule() is the main scheduler function.
4137 asmlinkage void __sched schedule(void)
4139 struct task_struct *prev, *next;
4140 unsigned long *switch_count;
4142 int cpu, hrtick = sched_feat(HRTICK);
4146 cpu = smp_processor_id();
4150 switch_count = &prev->nivcsw;
4152 release_kernel_lock(prev);
4153 need_resched_nonpreemptible:
4155 schedule_debug(prev);
4161 * Do the rq-clock update outside the rq lock:
4163 local_irq_disable();
4164 update_rq_clock(rq);
4165 spin_lock(&rq->lock);
4166 clear_tsk_need_resched(prev);
4168 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4169 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
4170 signal_pending(prev))) {
4171 prev->state = TASK_RUNNING;
4173 deactivate_task(rq, prev, 1);
4175 switch_count = &prev->nvcsw;
4179 if (prev->sched_class->pre_schedule)
4180 prev->sched_class->pre_schedule(rq, prev);
4183 if (unlikely(!rq->nr_running))
4184 idle_balance(cpu, rq);
4186 prev->sched_class->put_prev_task(rq, prev);
4187 next = pick_next_task(rq, prev);
4189 if (likely(prev != next)) {
4190 sched_info_switch(prev, next);
4196 context_switch(rq, prev, next); /* unlocks the rq */
4198 * the context switch might have flipped the stack from under
4199 * us, hence refresh the local variables.
4201 cpu = smp_processor_id();
4204 spin_unlock_irq(&rq->lock);
4209 if (unlikely(reacquire_kernel_lock(current) < 0))
4210 goto need_resched_nonpreemptible;
4212 preempt_enable_no_resched();
4213 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4216 EXPORT_SYMBOL(schedule);
4218 #ifdef CONFIG_PREEMPT
4220 * this is the entry point to schedule() from in-kernel preemption
4221 * off of preempt_enable. Kernel preemptions off return from interrupt
4222 * occur there and call schedule directly.
4224 asmlinkage void __sched preempt_schedule(void)
4226 struct thread_info *ti = current_thread_info();
4229 * If there is a non-zero preempt_count or interrupts are disabled,
4230 * we do not want to preempt the current task. Just return..
4232 if (likely(ti->preempt_count || irqs_disabled()))
4236 add_preempt_count(PREEMPT_ACTIVE);
4238 sub_preempt_count(PREEMPT_ACTIVE);
4241 * Check again in case we missed a preemption opportunity
4242 * between schedule and now.
4245 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4247 EXPORT_SYMBOL(preempt_schedule);
4250 * this is the entry point to schedule() from kernel preemption
4251 * off of irq context.
4252 * Note, that this is called and return with irqs disabled. This will
4253 * protect us against recursive calling from irq.
4255 asmlinkage void __sched preempt_schedule_irq(void)
4257 struct thread_info *ti = current_thread_info();
4259 /* Catch callers which need to be fixed */
4260 BUG_ON(ti->preempt_count || !irqs_disabled());
4263 add_preempt_count(PREEMPT_ACTIVE);
4266 local_irq_disable();
4267 sub_preempt_count(PREEMPT_ACTIVE);
4270 * Check again in case we missed a preemption opportunity
4271 * between schedule and now.
4274 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4277 #endif /* CONFIG_PREEMPT */
4279 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4282 return try_to_wake_up(curr->private, mode, sync);
4284 EXPORT_SYMBOL(default_wake_function);
4287 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4288 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4289 * number) then we wake all the non-exclusive tasks and one exclusive task.
4291 * There are circumstances in which we can try to wake a task which has already
4292 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4293 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4295 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4296 int nr_exclusive, int sync, void *key)
4298 wait_queue_t *curr, *next;
4300 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4301 unsigned flags = curr->flags;
4303 if (curr->func(curr, mode, sync, key) &&
4304 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4310 * __wake_up - wake up threads blocked on a waitqueue.
4312 * @mode: which threads
4313 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4314 * @key: is directly passed to the wakeup function
4316 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4317 int nr_exclusive, void *key)
4319 unsigned long flags;
4321 spin_lock_irqsave(&q->lock, flags);
4322 __wake_up_common(q, mode, nr_exclusive, 0, key);
4323 spin_unlock_irqrestore(&q->lock, flags);
4325 EXPORT_SYMBOL(__wake_up);
4328 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4330 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4332 __wake_up_common(q, mode, 1, 0, NULL);
4336 * __wake_up_sync - wake up threads blocked on a waitqueue.
4338 * @mode: which threads
4339 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4341 * The sync wakeup differs that the waker knows that it will schedule
4342 * away soon, so while the target thread will be woken up, it will not
4343 * be migrated to another CPU - ie. the two threads are 'synchronized'
4344 * with each other. This can prevent needless bouncing between CPUs.
4346 * On UP it can prevent extra preemption.
4349 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4351 unsigned long flags;
4357 if (unlikely(!nr_exclusive))
4360 spin_lock_irqsave(&q->lock, flags);
4361 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4362 spin_unlock_irqrestore(&q->lock, flags);
4364 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4366 void complete(struct completion *x)
4368 unsigned long flags;
4370 spin_lock_irqsave(&x->wait.lock, flags);
4372 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4373 spin_unlock_irqrestore(&x->wait.lock, flags);
4375 EXPORT_SYMBOL(complete);
4377 void complete_all(struct completion *x)
4379 unsigned long flags;
4381 spin_lock_irqsave(&x->wait.lock, flags);
4382 x->done += UINT_MAX/2;
4383 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4384 spin_unlock_irqrestore(&x->wait.lock, flags);
4386 EXPORT_SYMBOL(complete_all);
4388 static inline long __sched
4389 do_wait_for_common(struct completion *x, long timeout, int state)
4392 DECLARE_WAITQUEUE(wait, current);
4394 wait.flags |= WQ_FLAG_EXCLUSIVE;
4395 __add_wait_queue_tail(&x->wait, &wait);
4397 if ((state == TASK_INTERRUPTIBLE &&
4398 signal_pending(current)) ||
4399 (state == TASK_KILLABLE &&
4400 fatal_signal_pending(current))) {
4401 __remove_wait_queue(&x->wait, &wait);
4402 return -ERESTARTSYS;
4404 __set_current_state(state);
4405 spin_unlock_irq(&x->wait.lock);
4406 timeout = schedule_timeout(timeout);
4407 spin_lock_irq(&x->wait.lock);
4409 __remove_wait_queue(&x->wait, &wait);
4413 __remove_wait_queue(&x->wait, &wait);
4420 wait_for_common(struct completion *x, long timeout, int state)
4424 spin_lock_irq(&x->wait.lock);
4425 timeout = do_wait_for_common(x, timeout, state);
4426 spin_unlock_irq(&x->wait.lock);
4430 void __sched wait_for_completion(struct completion *x)
4432 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4434 EXPORT_SYMBOL(wait_for_completion);
4436 unsigned long __sched
4437 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4439 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4441 EXPORT_SYMBOL(wait_for_completion_timeout);
4443 int __sched wait_for_completion_interruptible(struct completion *x)
4445 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4446 if (t == -ERESTARTSYS)
4450 EXPORT_SYMBOL(wait_for_completion_interruptible);
4452 unsigned long __sched
4453 wait_for_completion_interruptible_timeout(struct completion *x,
4454 unsigned long timeout)
4456 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4458 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4460 int __sched wait_for_completion_killable(struct completion *x)
4462 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4463 if (t == -ERESTARTSYS)
4467 EXPORT_SYMBOL(wait_for_completion_killable);
4470 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4472 unsigned long flags;
4475 init_waitqueue_entry(&wait, current);
4477 __set_current_state(state);
4479 spin_lock_irqsave(&q->lock, flags);
4480 __add_wait_queue(q, &wait);
4481 spin_unlock(&q->lock);
4482 timeout = schedule_timeout(timeout);
4483 spin_lock_irq(&q->lock);
4484 __remove_wait_queue(q, &wait);
4485 spin_unlock_irqrestore(&q->lock, flags);
4490 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4492 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4494 EXPORT_SYMBOL(interruptible_sleep_on);
4497 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4499 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4501 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4503 void __sched sleep_on(wait_queue_head_t *q)
4505 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4507 EXPORT_SYMBOL(sleep_on);
4509 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4511 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4513 EXPORT_SYMBOL(sleep_on_timeout);
4515 #ifdef CONFIG_RT_MUTEXES
4518 * rt_mutex_setprio - set the current priority of a task
4520 * @prio: prio value (kernel-internal form)
4522 * This function changes the 'effective' priority of a task. It does
4523 * not touch ->normal_prio like __setscheduler().
4525 * Used by the rt_mutex code to implement priority inheritance logic.
4527 void rt_mutex_setprio(struct task_struct *p, int prio)
4529 unsigned long flags;
4530 int oldprio, on_rq, running;
4532 const struct sched_class *prev_class = p->sched_class;
4534 BUG_ON(prio < 0 || prio > MAX_PRIO);
4536 rq = task_rq_lock(p, &flags);
4537 update_rq_clock(rq);
4540 on_rq = p->se.on_rq;
4541 running = task_current(rq, p);
4543 dequeue_task(rq, p, 0);
4545 p->sched_class->put_prev_task(rq, p);
4548 p->sched_class = &rt_sched_class;
4550 p->sched_class = &fair_sched_class;
4555 p->sched_class->set_curr_task(rq);
4557 enqueue_task(rq, p, 0);
4559 check_class_changed(rq, p, prev_class, oldprio, running);
4561 task_rq_unlock(rq, &flags);
4566 void set_user_nice(struct task_struct *p, long nice)
4568 int old_prio, delta, on_rq;
4569 unsigned long flags;
4572 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4575 * We have to be careful, if called from sys_setpriority(),
4576 * the task might be in the middle of scheduling on another CPU.
4578 rq = task_rq_lock(p, &flags);
4579 update_rq_clock(rq);
4581 * The RT priorities are set via sched_setscheduler(), but we still
4582 * allow the 'normal' nice value to be set - but as expected
4583 * it wont have any effect on scheduling until the task is
4584 * SCHED_FIFO/SCHED_RR:
4586 if (task_has_rt_policy(p)) {
4587 p->static_prio = NICE_TO_PRIO(nice);
4590 on_rq = p->se.on_rq;
4592 dequeue_task(rq, p, 0);
4596 p->static_prio = NICE_TO_PRIO(nice);
4599 p->prio = effective_prio(p);
4600 delta = p->prio - old_prio;
4603 enqueue_task(rq, p, 0);
4606 * If the task increased its priority or is running and
4607 * lowered its priority, then reschedule its CPU:
4609 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4610 resched_task(rq->curr);
4613 task_rq_unlock(rq, &flags);
4615 EXPORT_SYMBOL(set_user_nice);
4618 * can_nice - check if a task can reduce its nice value
4622 int can_nice(const struct task_struct *p, const int nice)
4624 /* convert nice value [19,-20] to rlimit style value [1,40] */
4625 int nice_rlim = 20 - nice;
4627 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4628 capable(CAP_SYS_NICE));
4631 #ifdef __ARCH_WANT_SYS_NICE
4634 * sys_nice - change the priority of the current process.
4635 * @increment: priority increment
4637 * sys_setpriority is a more generic, but much slower function that
4638 * does similar things.
4640 asmlinkage long sys_nice(int increment)
4645 * Setpriority might change our priority at the same moment.
4646 * We don't have to worry. Conceptually one call occurs first
4647 * and we have a single winner.
4649 if (increment < -40)
4654 nice = PRIO_TO_NICE(current->static_prio) + increment;
4660 if (increment < 0 && !can_nice(current, nice))
4663 retval = security_task_setnice(current, nice);
4667 set_user_nice(current, nice);
4674 * task_prio - return the priority value of a given task.
4675 * @p: the task in question.
4677 * This is the priority value as seen by users in /proc.
4678 * RT tasks are offset by -200. Normal tasks are centered
4679 * around 0, value goes from -16 to +15.
4681 int task_prio(const struct task_struct *p)
4683 return p->prio - MAX_RT_PRIO;
4687 * task_nice - return the nice value of a given task.
4688 * @p: the task in question.
4690 int task_nice(const struct task_struct *p)
4692 return TASK_NICE(p);
4694 EXPORT_SYMBOL(task_nice);
4697 * idle_cpu - is a given cpu idle currently?
4698 * @cpu: the processor in question.
4700 int idle_cpu(int cpu)
4702 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4706 * idle_task - return the idle task for a given cpu.
4707 * @cpu: the processor in question.
4709 struct task_struct *idle_task(int cpu)
4711 return cpu_rq(cpu)->idle;
4715 * find_process_by_pid - find a process with a matching PID value.
4716 * @pid: the pid in question.
4718 static struct task_struct *find_process_by_pid(pid_t pid)
4720 return pid ? find_task_by_vpid(pid) : current;
4723 /* Actually do priority change: must hold rq lock. */
4725 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4727 BUG_ON(p->se.on_rq);
4730 switch (p->policy) {
4734 p->sched_class = &fair_sched_class;
4738 p->sched_class = &rt_sched_class;
4742 p->rt_priority = prio;
4743 p->normal_prio = normal_prio(p);
4744 /* we are holding p->pi_lock already */
4745 p->prio = rt_mutex_getprio(p);
4750 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4751 * @p: the task in question.
4752 * @policy: new policy.
4753 * @param: structure containing the new RT priority.
4755 * NOTE that the task may be already dead.
4757 int sched_setscheduler(struct task_struct *p, int policy,
4758 struct sched_param *param)
4760 int retval, oldprio, oldpolicy = -1, on_rq, running;
4761 unsigned long flags;
4762 const struct sched_class *prev_class = p->sched_class;
4765 /* may grab non-irq protected spin_locks */
4766 BUG_ON(in_interrupt());
4768 /* double check policy once rq lock held */
4770 policy = oldpolicy = p->policy;
4771 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4772 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4773 policy != SCHED_IDLE)
4776 * Valid priorities for SCHED_FIFO and SCHED_RR are
4777 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4778 * SCHED_BATCH and SCHED_IDLE is 0.
4780 if (param->sched_priority < 0 ||
4781 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4782 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4784 if (rt_policy(policy) != (param->sched_priority != 0))
4788 * Allow unprivileged RT tasks to decrease priority:
4790 if (!capable(CAP_SYS_NICE)) {
4791 if (rt_policy(policy)) {
4792 unsigned long rlim_rtprio;
4794 if (!lock_task_sighand(p, &flags))
4796 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4797 unlock_task_sighand(p, &flags);
4799 /* can't set/change the rt policy */
4800 if (policy != p->policy && !rlim_rtprio)
4803 /* can't increase priority */
4804 if (param->sched_priority > p->rt_priority &&
4805 param->sched_priority > rlim_rtprio)
4809 * Like positive nice levels, dont allow tasks to
4810 * move out of SCHED_IDLE either:
4812 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4815 /* can't change other user's priorities */
4816 if ((current->euid != p->euid) &&
4817 (current->euid != p->uid))
4821 #ifdef CONFIG_RT_GROUP_SCHED
4823 * Do not allow realtime tasks into groups that have no runtime
4826 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
4830 retval = security_task_setscheduler(p, policy, param);
4834 * make sure no PI-waiters arrive (or leave) while we are
4835 * changing the priority of the task:
4837 spin_lock_irqsave(&p->pi_lock, flags);
4839 * To be able to change p->policy safely, the apropriate
4840 * runqueue lock must be held.
4842 rq = __task_rq_lock(p);
4843 /* recheck policy now with rq lock held */
4844 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4845 policy = oldpolicy = -1;
4846 __task_rq_unlock(rq);
4847 spin_unlock_irqrestore(&p->pi_lock, flags);
4850 update_rq_clock(rq);
4851 on_rq = p->se.on_rq;
4852 running = task_current(rq, p);
4854 deactivate_task(rq, p, 0);
4856 p->sched_class->put_prev_task(rq, p);
4859 __setscheduler(rq, p, policy, param->sched_priority);
4862 p->sched_class->set_curr_task(rq);
4864 activate_task(rq, p, 0);
4866 check_class_changed(rq, p, prev_class, oldprio, running);
4868 __task_rq_unlock(rq);
4869 spin_unlock_irqrestore(&p->pi_lock, flags);
4871 rt_mutex_adjust_pi(p);
4875 EXPORT_SYMBOL_GPL(sched_setscheduler);
4878 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4880 struct sched_param lparam;
4881 struct task_struct *p;
4884 if (!param || pid < 0)
4886 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4891 p = find_process_by_pid(pid);
4893 retval = sched_setscheduler(p, policy, &lparam);
4900 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4901 * @pid: the pid in question.
4902 * @policy: new policy.
4903 * @param: structure containing the new RT priority.
4906 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4908 /* negative values for policy are not valid */
4912 return do_sched_setscheduler(pid, policy, param);
4916 * sys_sched_setparam - set/change the RT priority of a thread
4917 * @pid: the pid in question.
4918 * @param: structure containing the new RT priority.
4920 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4922 return do_sched_setscheduler(pid, -1, param);
4926 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4927 * @pid: the pid in question.
4929 asmlinkage long sys_sched_getscheduler(pid_t pid)
4931 struct task_struct *p;
4938 read_lock(&tasklist_lock);
4939 p = find_process_by_pid(pid);
4941 retval = security_task_getscheduler(p);
4945 read_unlock(&tasklist_lock);
4950 * sys_sched_getscheduler - get the RT priority of a thread
4951 * @pid: the pid in question.
4952 * @param: structure containing the RT priority.
4954 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4956 struct sched_param lp;
4957 struct task_struct *p;
4960 if (!param || pid < 0)
4963 read_lock(&tasklist_lock);
4964 p = find_process_by_pid(pid);
4969 retval = security_task_getscheduler(p);
4973 lp.sched_priority = p->rt_priority;
4974 read_unlock(&tasklist_lock);
4977 * This one might sleep, we cannot do it with a spinlock held ...
4979 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4984 read_unlock(&tasklist_lock);
4988 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
4990 cpumask_t cpus_allowed;
4991 cpumask_t new_mask = *in_mask;
4992 struct task_struct *p;
4996 read_lock(&tasklist_lock);
4998 p = find_process_by_pid(pid);
5000 read_unlock(&tasklist_lock);
5006 * It is not safe to call set_cpus_allowed with the
5007 * tasklist_lock held. We will bump the task_struct's
5008 * usage count and then drop tasklist_lock.
5011 read_unlock(&tasklist_lock);
5014 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5015 !capable(CAP_SYS_NICE))
5018 retval = security_task_setscheduler(p, 0, NULL);
5022 cpuset_cpus_allowed(p, &cpus_allowed);
5023 cpus_and(new_mask, new_mask, cpus_allowed);
5025 retval = set_cpus_allowed_ptr(p, &new_mask);
5028 cpuset_cpus_allowed(p, &cpus_allowed);
5029 if (!cpus_subset(new_mask, cpus_allowed)) {
5031 * We must have raced with a concurrent cpuset
5032 * update. Just reset the cpus_allowed to the
5033 * cpuset's cpus_allowed
5035 new_mask = cpus_allowed;
5045 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5046 cpumask_t *new_mask)
5048 if (len < sizeof(cpumask_t)) {
5049 memset(new_mask, 0, sizeof(cpumask_t));
5050 } else if (len > sizeof(cpumask_t)) {
5051 len = sizeof(cpumask_t);
5053 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5057 * sys_sched_setaffinity - set the cpu affinity of a process
5058 * @pid: pid of the process
5059 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5060 * @user_mask_ptr: user-space pointer to the new cpu mask
5062 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5063 unsigned long __user *user_mask_ptr)
5068 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5072 return sched_setaffinity(pid, &new_mask);
5076 * Represents all cpu's present in the system
5077 * In systems capable of hotplug, this map could dynamically grow
5078 * as new cpu's are detected in the system via any platform specific
5079 * method, such as ACPI for e.g.
5082 cpumask_t cpu_present_map __read_mostly;
5083 EXPORT_SYMBOL(cpu_present_map);
5086 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
5087 EXPORT_SYMBOL(cpu_online_map);
5089 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
5090 EXPORT_SYMBOL(cpu_possible_map);
5093 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5095 struct task_struct *p;
5099 read_lock(&tasklist_lock);
5102 p = find_process_by_pid(pid);
5106 retval = security_task_getscheduler(p);
5110 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5113 read_unlock(&tasklist_lock);
5120 * sys_sched_getaffinity - get the cpu affinity of a process
5121 * @pid: pid of the process
5122 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5123 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5125 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5126 unsigned long __user *user_mask_ptr)
5131 if (len < sizeof(cpumask_t))
5134 ret = sched_getaffinity(pid, &mask);
5138 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5141 return sizeof(cpumask_t);
5145 * sys_sched_yield - yield the current processor to other threads.
5147 * This function yields the current CPU to other tasks. If there are no
5148 * other threads running on this CPU then this function will return.
5150 asmlinkage long sys_sched_yield(void)
5152 struct rq *rq = this_rq_lock();
5154 schedstat_inc(rq, yld_count);
5155 current->sched_class->yield_task(rq);
5158 * Since we are going to call schedule() anyway, there's
5159 * no need to preempt or enable interrupts:
5161 __release(rq->lock);
5162 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5163 _raw_spin_unlock(&rq->lock);
5164 preempt_enable_no_resched();
5171 static void __cond_resched(void)
5173 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5174 __might_sleep(__FILE__, __LINE__);
5177 * The BKS might be reacquired before we have dropped
5178 * PREEMPT_ACTIVE, which could trigger a second
5179 * cond_resched() call.
5182 add_preempt_count(PREEMPT_ACTIVE);
5184 sub_preempt_count(PREEMPT_ACTIVE);
5185 } while (need_resched());
5188 int __sched _cond_resched(void)
5190 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5191 system_state == SYSTEM_RUNNING) {
5197 EXPORT_SYMBOL(_cond_resched);
5200 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5201 * call schedule, and on return reacquire the lock.
5203 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5204 * operations here to prevent schedule() from being called twice (once via
5205 * spin_unlock(), once by hand).
5207 int cond_resched_lock(spinlock_t *lock)
5209 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5212 if (spin_needbreak(lock) || resched) {
5214 if (resched && need_resched())
5223 EXPORT_SYMBOL(cond_resched_lock);
5225 int __sched cond_resched_softirq(void)
5227 BUG_ON(!in_softirq());
5229 if (need_resched() && system_state == SYSTEM_RUNNING) {
5237 EXPORT_SYMBOL(cond_resched_softirq);
5240 * yield - yield the current processor to other threads.
5242 * This is a shortcut for kernel-space yielding - it marks the
5243 * thread runnable and calls sys_sched_yield().
5245 void __sched yield(void)
5247 set_current_state(TASK_RUNNING);
5250 EXPORT_SYMBOL(yield);
5253 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5254 * that process accounting knows that this is a task in IO wait state.
5256 * But don't do that if it is a deliberate, throttling IO wait (this task
5257 * has set its backing_dev_info: the queue against which it should throttle)
5259 void __sched io_schedule(void)
5261 struct rq *rq = &__raw_get_cpu_var(runqueues);
5263 delayacct_blkio_start();
5264 atomic_inc(&rq->nr_iowait);
5266 atomic_dec(&rq->nr_iowait);
5267 delayacct_blkio_end();
5269 EXPORT_SYMBOL(io_schedule);
5271 long __sched io_schedule_timeout(long timeout)
5273 struct rq *rq = &__raw_get_cpu_var(runqueues);
5276 delayacct_blkio_start();
5277 atomic_inc(&rq->nr_iowait);
5278 ret = schedule_timeout(timeout);
5279 atomic_dec(&rq->nr_iowait);
5280 delayacct_blkio_end();
5285 * sys_sched_get_priority_max - return maximum RT priority.
5286 * @policy: scheduling class.
5288 * this syscall returns the maximum rt_priority that can be used
5289 * by a given scheduling class.
5291 asmlinkage long sys_sched_get_priority_max(int policy)
5298 ret = MAX_USER_RT_PRIO-1;
5310 * sys_sched_get_priority_min - return minimum RT priority.
5311 * @policy: scheduling class.
5313 * this syscall returns the minimum rt_priority that can be used
5314 * by a given scheduling class.
5316 asmlinkage long sys_sched_get_priority_min(int policy)
5334 * sys_sched_rr_get_interval - return the default timeslice of a process.
5335 * @pid: pid of the process.
5336 * @interval: userspace pointer to the timeslice value.
5338 * this syscall writes the default timeslice value of a given process
5339 * into the user-space timespec buffer. A value of '0' means infinity.
5342 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5344 struct task_struct *p;
5345 unsigned int time_slice;
5353 read_lock(&tasklist_lock);
5354 p = find_process_by_pid(pid);
5358 retval = security_task_getscheduler(p);
5363 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5364 * tasks that are on an otherwise idle runqueue:
5367 if (p->policy == SCHED_RR) {
5368 time_slice = DEF_TIMESLICE;
5369 } else if (p->policy != SCHED_FIFO) {
5370 struct sched_entity *se = &p->se;
5371 unsigned long flags;
5374 rq = task_rq_lock(p, &flags);
5375 if (rq->cfs.load.weight)
5376 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5377 task_rq_unlock(rq, &flags);
5379 read_unlock(&tasklist_lock);
5380 jiffies_to_timespec(time_slice, &t);
5381 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5385 read_unlock(&tasklist_lock);
5389 static const char stat_nam[] = "RSDTtZX";
5391 void sched_show_task(struct task_struct *p)
5393 unsigned long free = 0;
5396 state = p->state ? __ffs(p->state) + 1 : 0;
5397 printk(KERN_INFO "%-13.13s %c", p->comm,
5398 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5399 #if BITS_PER_LONG == 32
5400 if (state == TASK_RUNNING)
5401 printk(KERN_CONT " running ");
5403 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5405 if (state == TASK_RUNNING)
5406 printk(KERN_CONT " running task ");
5408 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5410 #ifdef CONFIG_DEBUG_STACK_USAGE
5412 unsigned long *n = end_of_stack(p);
5415 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5418 printk(KERN_CONT "%5lu %5d %6d\n", free,
5419 task_pid_nr(p), task_pid_nr(p->real_parent));
5421 show_stack(p, NULL);
5424 void show_state_filter(unsigned long state_filter)
5426 struct task_struct *g, *p;
5428 #if BITS_PER_LONG == 32
5430 " task PC stack pid father\n");
5433 " task PC stack pid father\n");
5435 read_lock(&tasklist_lock);
5436 do_each_thread(g, p) {
5438 * reset the NMI-timeout, listing all files on a slow
5439 * console might take alot of time:
5441 touch_nmi_watchdog();
5442 if (!state_filter || (p->state & state_filter))
5444 } while_each_thread(g, p);
5446 touch_all_softlockup_watchdogs();
5448 #ifdef CONFIG_SCHED_DEBUG
5449 sysrq_sched_debug_show();
5451 read_unlock(&tasklist_lock);
5453 * Only show locks if all tasks are dumped:
5455 if (state_filter == -1)
5456 debug_show_all_locks();
5459 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5461 idle->sched_class = &idle_sched_class;
5465 * init_idle - set up an idle thread for a given CPU
5466 * @idle: task in question
5467 * @cpu: cpu the idle task belongs to
5469 * NOTE: this function does not set the idle thread's NEED_RESCHED
5470 * flag, to make booting more robust.
5472 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5474 struct rq *rq = cpu_rq(cpu);
5475 unsigned long flags;
5478 idle->se.exec_start = sched_clock();
5480 idle->prio = idle->normal_prio = MAX_PRIO;
5481 idle->cpus_allowed = cpumask_of_cpu(cpu);
5482 __set_task_cpu(idle, cpu);
5484 spin_lock_irqsave(&rq->lock, flags);
5485 rq->curr = rq->idle = idle;
5486 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5489 spin_unlock_irqrestore(&rq->lock, flags);
5491 /* Set the preempt count _outside_ the spinlocks! */
5492 #if defined(CONFIG_PREEMPT)
5493 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5495 task_thread_info(idle)->preempt_count = 0;
5498 * The idle tasks have their own, simple scheduling class:
5500 idle->sched_class = &idle_sched_class;
5504 * In a system that switches off the HZ timer nohz_cpu_mask
5505 * indicates which cpus entered this state. This is used
5506 * in the rcu update to wait only for active cpus. For system
5507 * which do not switch off the HZ timer nohz_cpu_mask should
5508 * always be CPU_MASK_NONE.
5510 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5513 * Increase the granularity value when there are more CPUs,
5514 * because with more CPUs the 'effective latency' as visible
5515 * to users decreases. But the relationship is not linear,
5516 * so pick a second-best guess by going with the log2 of the
5519 * This idea comes from the SD scheduler of Con Kolivas:
5521 static inline void sched_init_granularity(void)
5523 unsigned int factor = 1 + ilog2(num_online_cpus());
5524 const unsigned long limit = 200000000;
5526 sysctl_sched_min_granularity *= factor;
5527 if (sysctl_sched_min_granularity > limit)
5528 sysctl_sched_min_granularity = limit;
5530 sysctl_sched_latency *= factor;
5531 if (sysctl_sched_latency > limit)
5532 sysctl_sched_latency = limit;
5534 sysctl_sched_wakeup_granularity *= factor;
5539 * This is how migration works:
5541 * 1) we queue a struct migration_req structure in the source CPU's
5542 * runqueue and wake up that CPU's migration thread.
5543 * 2) we down() the locked semaphore => thread blocks.
5544 * 3) migration thread wakes up (implicitly it forces the migrated
5545 * thread off the CPU)
5546 * 4) it gets the migration request and checks whether the migrated
5547 * task is still in the wrong runqueue.
5548 * 5) if it's in the wrong runqueue then the migration thread removes
5549 * it and puts it into the right queue.
5550 * 6) migration thread up()s the semaphore.
5551 * 7) we wake up and the migration is done.
5555 * Change a given task's CPU affinity. Migrate the thread to a
5556 * proper CPU and schedule it away if the CPU it's executing on
5557 * is removed from the allowed bitmask.
5559 * NOTE: the caller must have a valid reference to the task, the
5560 * task must not exit() & deallocate itself prematurely. The
5561 * call is not atomic; no spinlocks may be held.
5563 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5565 struct migration_req req;
5566 unsigned long flags;
5570 rq = task_rq_lock(p, &flags);
5571 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5576 if (p->sched_class->set_cpus_allowed)
5577 p->sched_class->set_cpus_allowed(p, new_mask);
5579 p->cpus_allowed = *new_mask;
5580 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5583 /* Can the task run on the task's current CPU? If so, we're done */
5584 if (cpu_isset(task_cpu(p), *new_mask))
5587 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5588 /* Need help from migration thread: drop lock and wait. */
5589 task_rq_unlock(rq, &flags);
5590 wake_up_process(rq->migration_thread);
5591 wait_for_completion(&req.done);
5592 tlb_migrate_finish(p->mm);
5596 task_rq_unlock(rq, &flags);
5600 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5603 * Move (not current) task off this cpu, onto dest cpu. We're doing
5604 * this because either it can't run here any more (set_cpus_allowed()
5605 * away from this CPU, or CPU going down), or because we're
5606 * attempting to rebalance this task on exec (sched_exec).
5608 * So we race with normal scheduler movements, but that's OK, as long
5609 * as the task is no longer on this CPU.
5611 * Returns non-zero if task was successfully migrated.
5613 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5615 struct rq *rq_dest, *rq_src;
5618 if (unlikely(cpu_is_offline(dest_cpu)))
5621 rq_src = cpu_rq(src_cpu);
5622 rq_dest = cpu_rq(dest_cpu);
5624 double_rq_lock(rq_src, rq_dest);
5625 /* Already moved. */
5626 if (task_cpu(p) != src_cpu)
5628 /* Affinity changed (again). */
5629 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5632 on_rq = p->se.on_rq;
5634 deactivate_task(rq_src, p, 0);
5636 set_task_cpu(p, dest_cpu);
5638 activate_task(rq_dest, p, 0);
5639 check_preempt_curr(rq_dest, p);
5643 double_rq_unlock(rq_src, rq_dest);
5648 * migration_thread - this is a highprio system thread that performs
5649 * thread migration by bumping thread off CPU then 'pushing' onto
5652 static int migration_thread(void *data)
5654 int cpu = (long)data;
5658 BUG_ON(rq->migration_thread != current);
5660 set_current_state(TASK_INTERRUPTIBLE);
5661 while (!kthread_should_stop()) {
5662 struct migration_req *req;
5663 struct list_head *head;
5665 spin_lock_irq(&rq->lock);
5667 if (cpu_is_offline(cpu)) {
5668 spin_unlock_irq(&rq->lock);
5672 if (rq->active_balance) {
5673 active_load_balance(rq, cpu);
5674 rq->active_balance = 0;
5677 head = &rq->migration_queue;
5679 if (list_empty(head)) {
5680 spin_unlock_irq(&rq->lock);
5682 set_current_state(TASK_INTERRUPTIBLE);
5685 req = list_entry(head->next, struct migration_req, list);
5686 list_del_init(head->next);
5688 spin_unlock(&rq->lock);
5689 __migrate_task(req->task, cpu, req->dest_cpu);
5692 complete(&req->done);
5694 __set_current_state(TASK_RUNNING);
5698 /* Wait for kthread_stop */
5699 set_current_state(TASK_INTERRUPTIBLE);
5700 while (!kthread_should_stop()) {
5702 set_current_state(TASK_INTERRUPTIBLE);
5704 __set_current_state(TASK_RUNNING);
5708 #ifdef CONFIG_HOTPLUG_CPU
5710 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5714 local_irq_disable();
5715 ret = __migrate_task(p, src_cpu, dest_cpu);
5721 * Figure out where task on dead CPU should go, use force if necessary.
5722 * NOTE: interrupts should be disabled by the caller
5724 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5726 unsigned long flags;
5733 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5734 cpus_and(mask, mask, p->cpus_allowed);
5735 dest_cpu = any_online_cpu(mask);
5737 /* On any allowed CPU? */
5738 if (dest_cpu >= nr_cpu_ids)
5739 dest_cpu = any_online_cpu(p->cpus_allowed);
5741 /* No more Mr. Nice Guy. */
5742 if (dest_cpu >= nr_cpu_ids) {
5743 cpumask_t cpus_allowed;
5745 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5747 * Try to stay on the same cpuset, where the
5748 * current cpuset may be a subset of all cpus.
5749 * The cpuset_cpus_allowed_locked() variant of
5750 * cpuset_cpus_allowed() will not block. It must be
5751 * called within calls to cpuset_lock/cpuset_unlock.
5753 rq = task_rq_lock(p, &flags);
5754 p->cpus_allowed = cpus_allowed;
5755 dest_cpu = any_online_cpu(p->cpus_allowed);
5756 task_rq_unlock(rq, &flags);
5759 * Don't tell them about moving exiting tasks or
5760 * kernel threads (both mm NULL), since they never
5763 if (p->mm && printk_ratelimit()) {
5764 printk(KERN_INFO "process %d (%s) no "
5765 "longer affine to cpu%d\n",
5766 task_pid_nr(p), p->comm, dead_cpu);
5769 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5773 * While a dead CPU has no uninterruptible tasks queued at this point,
5774 * it might still have a nonzero ->nr_uninterruptible counter, because
5775 * for performance reasons the counter is not stricly tracking tasks to
5776 * their home CPUs. So we just add the counter to another CPU's counter,
5777 * to keep the global sum constant after CPU-down:
5779 static void migrate_nr_uninterruptible(struct rq *rq_src)
5781 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5782 unsigned long flags;
5784 local_irq_save(flags);
5785 double_rq_lock(rq_src, rq_dest);
5786 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5787 rq_src->nr_uninterruptible = 0;
5788 double_rq_unlock(rq_src, rq_dest);
5789 local_irq_restore(flags);
5792 /* Run through task list and migrate tasks from the dead cpu. */
5793 static void migrate_live_tasks(int src_cpu)
5795 struct task_struct *p, *t;
5797 read_lock(&tasklist_lock);
5799 do_each_thread(t, p) {
5803 if (task_cpu(p) == src_cpu)
5804 move_task_off_dead_cpu(src_cpu, p);
5805 } while_each_thread(t, p);
5807 read_unlock(&tasklist_lock);
5811 * Schedules idle task to be the next runnable task on current CPU.
5812 * It does so by boosting its priority to highest possible.
5813 * Used by CPU offline code.
5815 void sched_idle_next(void)
5817 int this_cpu = smp_processor_id();
5818 struct rq *rq = cpu_rq(this_cpu);
5819 struct task_struct *p = rq->idle;
5820 unsigned long flags;
5822 /* cpu has to be offline */
5823 BUG_ON(cpu_online(this_cpu));
5826 * Strictly not necessary since rest of the CPUs are stopped by now
5827 * and interrupts disabled on the current cpu.
5829 spin_lock_irqsave(&rq->lock, flags);
5831 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5833 update_rq_clock(rq);
5834 activate_task(rq, p, 0);
5836 spin_unlock_irqrestore(&rq->lock, flags);
5840 * Ensures that the idle task is using init_mm right before its cpu goes
5843 void idle_task_exit(void)
5845 struct mm_struct *mm = current->active_mm;
5847 BUG_ON(cpu_online(smp_processor_id()));
5850 switch_mm(mm, &init_mm, current);
5854 /* called under rq->lock with disabled interrupts */
5855 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5857 struct rq *rq = cpu_rq(dead_cpu);
5859 /* Must be exiting, otherwise would be on tasklist. */
5860 BUG_ON(!p->exit_state);
5862 /* Cannot have done final schedule yet: would have vanished. */
5863 BUG_ON(p->state == TASK_DEAD);
5868 * Drop lock around migration; if someone else moves it,
5869 * that's OK. No task can be added to this CPU, so iteration is
5872 spin_unlock_irq(&rq->lock);
5873 move_task_off_dead_cpu(dead_cpu, p);
5874 spin_lock_irq(&rq->lock);
5879 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5880 static void migrate_dead_tasks(unsigned int dead_cpu)
5882 struct rq *rq = cpu_rq(dead_cpu);
5883 struct task_struct *next;
5886 if (!rq->nr_running)
5888 update_rq_clock(rq);
5889 next = pick_next_task(rq, rq->curr);
5892 migrate_dead(dead_cpu, next);
5896 #endif /* CONFIG_HOTPLUG_CPU */
5898 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5900 static struct ctl_table sd_ctl_dir[] = {
5902 .procname = "sched_domain",
5908 static struct ctl_table sd_ctl_root[] = {
5910 .ctl_name = CTL_KERN,
5911 .procname = "kernel",
5913 .child = sd_ctl_dir,
5918 static struct ctl_table *sd_alloc_ctl_entry(int n)
5920 struct ctl_table *entry =
5921 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5926 static void sd_free_ctl_entry(struct ctl_table **tablep)
5928 struct ctl_table *entry;
5931 * In the intermediate directories, both the child directory and
5932 * procname are dynamically allocated and could fail but the mode
5933 * will always be set. In the lowest directory the names are
5934 * static strings and all have proc handlers.
5936 for (entry = *tablep; entry->mode; entry++) {
5938 sd_free_ctl_entry(&entry->child);
5939 if (entry->proc_handler == NULL)
5940 kfree(entry->procname);
5948 set_table_entry(struct ctl_table *entry,
5949 const char *procname, void *data, int maxlen,
5950 mode_t mode, proc_handler *proc_handler)
5952 entry->procname = procname;
5954 entry->maxlen = maxlen;
5956 entry->proc_handler = proc_handler;
5959 static struct ctl_table *
5960 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5962 struct ctl_table *table = sd_alloc_ctl_entry(12);
5967 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5968 sizeof(long), 0644, proc_doulongvec_minmax);
5969 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5970 sizeof(long), 0644, proc_doulongvec_minmax);
5971 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5972 sizeof(int), 0644, proc_dointvec_minmax);
5973 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5974 sizeof(int), 0644, proc_dointvec_minmax);
5975 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5976 sizeof(int), 0644, proc_dointvec_minmax);
5977 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5978 sizeof(int), 0644, proc_dointvec_minmax);
5979 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5980 sizeof(int), 0644, proc_dointvec_minmax);
5981 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5982 sizeof(int), 0644, proc_dointvec_minmax);
5983 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5984 sizeof(int), 0644, proc_dointvec_minmax);
5985 set_table_entry(&table[9], "cache_nice_tries",
5986 &sd->cache_nice_tries,
5987 sizeof(int), 0644, proc_dointvec_minmax);
5988 set_table_entry(&table[10], "flags", &sd->flags,
5989 sizeof(int), 0644, proc_dointvec_minmax);
5990 /* &table[11] is terminator */
5995 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5997 struct ctl_table *entry, *table;
5998 struct sched_domain *sd;
5999 int domain_num = 0, i;
6002 for_each_domain(cpu, sd)
6004 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6009 for_each_domain(cpu, sd) {
6010 snprintf(buf, 32, "domain%d", i);
6011 entry->procname = kstrdup(buf, GFP_KERNEL);
6013 entry->child = sd_alloc_ctl_domain_table(sd);
6020 static struct ctl_table_header *sd_sysctl_header;
6021 static void register_sched_domain_sysctl(void)
6023 int i, cpu_num = num_online_cpus();
6024 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6027 WARN_ON(sd_ctl_dir[0].child);
6028 sd_ctl_dir[0].child = entry;
6033 for_each_online_cpu(i) {
6034 snprintf(buf, 32, "cpu%d", i);
6035 entry->procname = kstrdup(buf, GFP_KERNEL);
6037 entry->child = sd_alloc_ctl_cpu_table(i);
6041 WARN_ON(sd_sysctl_header);
6042 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6045 /* may be called multiple times per register */
6046 static void unregister_sched_domain_sysctl(void)
6048 if (sd_sysctl_header)
6049 unregister_sysctl_table(sd_sysctl_header);
6050 sd_sysctl_header = NULL;
6051 if (sd_ctl_dir[0].child)
6052 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6055 static void register_sched_domain_sysctl(void)
6058 static void unregister_sched_domain_sysctl(void)
6064 * migration_call - callback that gets triggered when a CPU is added.
6065 * Here we can start up the necessary migration thread for the new CPU.
6067 static int __cpuinit
6068 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6070 struct task_struct *p;
6071 int cpu = (long)hcpu;
6072 unsigned long flags;
6077 case CPU_UP_PREPARE:
6078 case CPU_UP_PREPARE_FROZEN:
6079 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6082 kthread_bind(p, cpu);
6083 /* Must be high prio: stop_machine expects to yield to it. */
6084 rq = task_rq_lock(p, &flags);
6085 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6086 task_rq_unlock(rq, &flags);
6087 cpu_rq(cpu)->migration_thread = p;
6091 case CPU_ONLINE_FROZEN:
6092 /* Strictly unnecessary, as first user will wake it. */
6093 wake_up_process(cpu_rq(cpu)->migration_thread);
6095 /* Update our root-domain */
6097 spin_lock_irqsave(&rq->lock, flags);
6099 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6100 cpu_set(cpu, rq->rd->online);
6102 spin_unlock_irqrestore(&rq->lock, flags);
6105 #ifdef CONFIG_HOTPLUG_CPU
6106 case CPU_UP_CANCELED:
6107 case CPU_UP_CANCELED_FROZEN:
6108 if (!cpu_rq(cpu)->migration_thread)
6110 /* Unbind it from offline cpu so it can run. Fall thru. */
6111 kthread_bind(cpu_rq(cpu)->migration_thread,
6112 any_online_cpu(cpu_online_map));
6113 kthread_stop(cpu_rq(cpu)->migration_thread);
6114 cpu_rq(cpu)->migration_thread = NULL;
6118 case CPU_DEAD_FROZEN:
6119 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6120 migrate_live_tasks(cpu);
6122 kthread_stop(rq->migration_thread);
6123 rq->migration_thread = NULL;
6124 /* Idle task back to normal (off runqueue, low prio) */
6125 spin_lock_irq(&rq->lock);
6126 update_rq_clock(rq);
6127 deactivate_task(rq, rq->idle, 0);
6128 rq->idle->static_prio = MAX_PRIO;
6129 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6130 rq->idle->sched_class = &idle_sched_class;
6131 migrate_dead_tasks(cpu);
6132 spin_unlock_irq(&rq->lock);
6134 migrate_nr_uninterruptible(rq);
6135 BUG_ON(rq->nr_running != 0);
6138 * No need to migrate the tasks: it was best-effort if
6139 * they didn't take sched_hotcpu_mutex. Just wake up
6142 spin_lock_irq(&rq->lock);
6143 while (!list_empty(&rq->migration_queue)) {
6144 struct migration_req *req;
6146 req = list_entry(rq->migration_queue.next,
6147 struct migration_req, list);
6148 list_del_init(&req->list);
6149 complete(&req->done);
6151 spin_unlock_irq(&rq->lock);
6155 case CPU_DYING_FROZEN:
6156 /* Update our root-domain */
6158 spin_lock_irqsave(&rq->lock, flags);
6160 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6161 cpu_clear(cpu, rq->rd->online);
6163 spin_unlock_irqrestore(&rq->lock, flags);
6170 /* Register at highest priority so that task migration (migrate_all_tasks)
6171 * happens before everything else.
6173 static struct notifier_block __cpuinitdata migration_notifier = {
6174 .notifier_call = migration_call,
6178 void __init migration_init(void)
6180 void *cpu = (void *)(long)smp_processor_id();
6183 /* Start one for the boot CPU: */
6184 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6185 BUG_ON(err == NOTIFY_BAD);
6186 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6187 register_cpu_notifier(&migration_notifier);
6193 #ifdef CONFIG_SCHED_DEBUG
6195 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6196 cpumask_t *groupmask)
6198 struct sched_group *group = sd->groups;
6201 cpulist_scnprintf(str, sizeof(str), sd->span);
6202 cpus_clear(*groupmask);
6204 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6206 if (!(sd->flags & SD_LOAD_BALANCE)) {
6207 printk("does not load-balance\n");
6209 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6214 printk(KERN_CONT "span %s\n", str);
6216 if (!cpu_isset(cpu, sd->span)) {
6217 printk(KERN_ERR "ERROR: domain->span does not contain "
6220 if (!cpu_isset(cpu, group->cpumask)) {
6221 printk(KERN_ERR "ERROR: domain->groups does not contain"
6225 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6229 printk(KERN_ERR "ERROR: group is NULL\n");
6233 if (!group->__cpu_power) {
6234 printk(KERN_CONT "\n");
6235 printk(KERN_ERR "ERROR: domain->cpu_power not "
6240 if (!cpus_weight(group->cpumask)) {
6241 printk(KERN_CONT "\n");
6242 printk(KERN_ERR "ERROR: empty group\n");
6246 if (cpus_intersects(*groupmask, group->cpumask)) {
6247 printk(KERN_CONT "\n");
6248 printk(KERN_ERR "ERROR: repeated CPUs\n");
6252 cpus_or(*groupmask, *groupmask, group->cpumask);
6254 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6255 printk(KERN_CONT " %s", str);
6257 group = group->next;
6258 } while (group != sd->groups);
6259 printk(KERN_CONT "\n");
6261 if (!cpus_equal(sd->span, *groupmask))
6262 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6264 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6265 printk(KERN_ERR "ERROR: parent span is not a superset "
6266 "of domain->span\n");
6270 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6272 cpumask_t *groupmask;
6276 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6280 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6282 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6284 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6289 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6299 # define sched_domain_debug(sd, cpu) do { } while (0)
6302 static int sd_degenerate(struct sched_domain *sd)
6304 if (cpus_weight(sd->span) == 1)
6307 /* Following flags need at least 2 groups */
6308 if (sd->flags & (SD_LOAD_BALANCE |
6309 SD_BALANCE_NEWIDLE |
6313 SD_SHARE_PKG_RESOURCES)) {
6314 if (sd->groups != sd->groups->next)
6318 /* Following flags don't use groups */
6319 if (sd->flags & (SD_WAKE_IDLE |
6328 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6330 unsigned long cflags = sd->flags, pflags = parent->flags;
6332 if (sd_degenerate(parent))
6335 if (!cpus_equal(sd->span, parent->span))
6338 /* Does parent contain flags not in child? */
6339 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6340 if (cflags & SD_WAKE_AFFINE)
6341 pflags &= ~SD_WAKE_BALANCE;
6342 /* Flags needing groups don't count if only 1 group in parent */
6343 if (parent->groups == parent->groups->next) {
6344 pflags &= ~(SD_LOAD_BALANCE |
6345 SD_BALANCE_NEWIDLE |
6349 SD_SHARE_PKG_RESOURCES);
6351 if (~cflags & pflags)
6357 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6359 unsigned long flags;
6360 const struct sched_class *class;
6362 spin_lock_irqsave(&rq->lock, flags);
6365 struct root_domain *old_rd = rq->rd;
6367 for (class = sched_class_highest; class; class = class->next) {
6368 if (class->leave_domain)
6369 class->leave_domain(rq);
6372 cpu_clear(rq->cpu, old_rd->span);
6373 cpu_clear(rq->cpu, old_rd->online);
6375 if (atomic_dec_and_test(&old_rd->refcount))
6379 atomic_inc(&rd->refcount);
6382 cpu_set(rq->cpu, rd->span);
6383 if (cpu_isset(rq->cpu, cpu_online_map))
6384 cpu_set(rq->cpu, rd->online);
6386 for (class = sched_class_highest; class; class = class->next) {
6387 if (class->join_domain)
6388 class->join_domain(rq);
6391 spin_unlock_irqrestore(&rq->lock, flags);
6394 static void init_rootdomain(struct root_domain *rd)
6396 memset(rd, 0, sizeof(*rd));
6398 cpus_clear(rd->span);
6399 cpus_clear(rd->online);
6401 cpupri_init(&rd->cpupri);
6404 static void init_defrootdomain(void)
6406 init_rootdomain(&def_root_domain);
6407 atomic_set(&def_root_domain.refcount, 1);
6410 static struct root_domain *alloc_rootdomain(void)
6412 struct root_domain *rd;
6414 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6418 init_rootdomain(rd);
6424 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6425 * hold the hotplug lock.
6428 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6430 struct rq *rq = cpu_rq(cpu);
6431 struct sched_domain *tmp;
6433 /* Remove the sched domains which do not contribute to scheduling. */
6434 for (tmp = sd; tmp; tmp = tmp->parent) {
6435 struct sched_domain *parent = tmp->parent;
6438 if (sd_parent_degenerate(tmp, parent)) {
6439 tmp->parent = parent->parent;
6441 parent->parent->child = tmp;
6445 if (sd && sd_degenerate(sd)) {
6451 sched_domain_debug(sd, cpu);
6453 rq_attach_root(rq, rd);
6454 rcu_assign_pointer(rq->sd, sd);
6457 /* cpus with isolated domains */
6458 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6460 /* Setup the mask of cpus configured for isolated domains */
6461 static int __init isolated_cpu_setup(char *str)
6463 int ints[NR_CPUS], i;
6465 str = get_options(str, ARRAY_SIZE(ints), ints);
6466 cpus_clear(cpu_isolated_map);
6467 for (i = 1; i <= ints[0]; i++)
6468 if (ints[i] < NR_CPUS)
6469 cpu_set(ints[i], cpu_isolated_map);
6473 __setup("isolcpus=", isolated_cpu_setup);
6476 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6477 * to a function which identifies what group(along with sched group) a CPU
6478 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6479 * (due to the fact that we keep track of groups covered with a cpumask_t).
6481 * init_sched_build_groups will build a circular linked list of the groups
6482 * covered by the given span, and will set each group's ->cpumask correctly,
6483 * and ->cpu_power to 0.
6486 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6487 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6488 struct sched_group **sg,
6489 cpumask_t *tmpmask),
6490 cpumask_t *covered, cpumask_t *tmpmask)
6492 struct sched_group *first = NULL, *last = NULL;
6495 cpus_clear(*covered);
6497 for_each_cpu_mask(i, *span) {
6498 struct sched_group *sg;
6499 int group = group_fn(i, cpu_map, &sg, tmpmask);
6502 if (cpu_isset(i, *covered))
6505 cpus_clear(sg->cpumask);
6506 sg->__cpu_power = 0;
6508 for_each_cpu_mask(j, *span) {
6509 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6512 cpu_set(j, *covered);
6513 cpu_set(j, sg->cpumask);
6524 #define SD_NODES_PER_DOMAIN 16
6529 * find_next_best_node - find the next node to include in a sched_domain
6530 * @node: node whose sched_domain we're building
6531 * @used_nodes: nodes already in the sched_domain
6533 * Find the next node to include in a given scheduling domain. Simply
6534 * finds the closest node not already in the @used_nodes map.
6536 * Should use nodemask_t.
6538 static int find_next_best_node(int node, nodemask_t *used_nodes)
6540 int i, n, val, min_val, best_node = 0;
6544 for (i = 0; i < MAX_NUMNODES; i++) {
6545 /* Start at @node */
6546 n = (node + i) % MAX_NUMNODES;
6548 if (!nr_cpus_node(n))
6551 /* Skip already used nodes */
6552 if (node_isset(n, *used_nodes))
6555 /* Simple min distance search */
6556 val = node_distance(node, n);
6558 if (val < min_val) {
6564 node_set(best_node, *used_nodes);
6569 * sched_domain_node_span - get a cpumask for a node's sched_domain
6570 * @node: node whose cpumask we're constructing
6571 * @span: resulting cpumask
6573 * Given a node, construct a good cpumask for its sched_domain to span. It
6574 * should be one that prevents unnecessary balancing, but also spreads tasks
6577 static void sched_domain_node_span(int node, cpumask_t *span)
6579 nodemask_t used_nodes;
6580 node_to_cpumask_ptr(nodemask, node);
6584 nodes_clear(used_nodes);
6586 cpus_or(*span, *span, *nodemask);
6587 node_set(node, used_nodes);
6589 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6590 int next_node = find_next_best_node(node, &used_nodes);
6592 node_to_cpumask_ptr_next(nodemask, next_node);
6593 cpus_or(*span, *span, *nodemask);
6598 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6601 * SMT sched-domains:
6603 #ifdef CONFIG_SCHED_SMT
6604 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6605 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6608 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6612 *sg = &per_cpu(sched_group_cpus, cpu);
6618 * multi-core sched-domains:
6620 #ifdef CONFIG_SCHED_MC
6621 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6622 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6625 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6627 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6632 *mask = per_cpu(cpu_sibling_map, cpu);
6633 cpus_and(*mask, *mask, *cpu_map);
6634 group = first_cpu(*mask);
6636 *sg = &per_cpu(sched_group_core, group);
6639 #elif defined(CONFIG_SCHED_MC)
6641 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6645 *sg = &per_cpu(sched_group_core, cpu);
6650 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6651 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6654 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6658 #ifdef CONFIG_SCHED_MC
6659 *mask = cpu_coregroup_map(cpu);
6660 cpus_and(*mask, *mask, *cpu_map);
6661 group = first_cpu(*mask);
6662 #elif defined(CONFIG_SCHED_SMT)
6663 *mask = per_cpu(cpu_sibling_map, cpu);
6664 cpus_and(*mask, *mask, *cpu_map);
6665 group = first_cpu(*mask);
6670 *sg = &per_cpu(sched_group_phys, group);
6676 * The init_sched_build_groups can't handle what we want to do with node
6677 * groups, so roll our own. Now each node has its own list of groups which
6678 * gets dynamically allocated.
6680 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6681 static struct sched_group ***sched_group_nodes_bycpu;
6683 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6684 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6686 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6687 struct sched_group **sg, cpumask_t *nodemask)
6691 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6692 cpus_and(*nodemask, *nodemask, *cpu_map);
6693 group = first_cpu(*nodemask);
6696 *sg = &per_cpu(sched_group_allnodes, group);
6700 static void init_numa_sched_groups_power(struct sched_group *group_head)
6702 struct sched_group *sg = group_head;
6708 for_each_cpu_mask(j, sg->cpumask) {
6709 struct sched_domain *sd;
6711 sd = &per_cpu(phys_domains, j);
6712 if (j != first_cpu(sd->groups->cpumask)) {
6714 * Only add "power" once for each
6720 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6723 } while (sg != group_head);
6728 /* Free memory allocated for various sched_group structures */
6729 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6733 for_each_cpu_mask(cpu, *cpu_map) {
6734 struct sched_group **sched_group_nodes
6735 = sched_group_nodes_bycpu[cpu];
6737 if (!sched_group_nodes)
6740 for (i = 0; i < MAX_NUMNODES; i++) {
6741 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6743 *nodemask = node_to_cpumask(i);
6744 cpus_and(*nodemask, *nodemask, *cpu_map);
6745 if (cpus_empty(*nodemask))
6755 if (oldsg != sched_group_nodes[i])
6758 kfree(sched_group_nodes);
6759 sched_group_nodes_bycpu[cpu] = NULL;
6763 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6769 * Initialize sched groups cpu_power.
6771 * cpu_power indicates the capacity of sched group, which is used while
6772 * distributing the load between different sched groups in a sched domain.
6773 * Typically cpu_power for all the groups in a sched domain will be same unless
6774 * there are asymmetries in the topology. If there are asymmetries, group
6775 * having more cpu_power will pickup more load compared to the group having
6778 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6779 * the maximum number of tasks a group can handle in the presence of other idle
6780 * or lightly loaded groups in the same sched domain.
6782 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6784 struct sched_domain *child;
6785 struct sched_group *group;
6787 WARN_ON(!sd || !sd->groups);
6789 if (cpu != first_cpu(sd->groups->cpumask))
6794 sd->groups->__cpu_power = 0;
6797 * For perf policy, if the groups in child domain share resources
6798 * (for example cores sharing some portions of the cache hierarchy
6799 * or SMT), then set this domain groups cpu_power such that each group
6800 * can handle only one task, when there are other idle groups in the
6801 * same sched domain.
6803 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6805 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6806 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6811 * add cpu_power of each child group to this groups cpu_power
6813 group = child->groups;
6815 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6816 group = group->next;
6817 } while (group != child->groups);
6821 * Initializers for schedule domains
6822 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6825 #define SD_INIT(sd, type) sd_init_##type(sd)
6826 #define SD_INIT_FUNC(type) \
6827 static noinline void sd_init_##type(struct sched_domain *sd) \
6829 memset(sd, 0, sizeof(*sd)); \
6830 *sd = SD_##type##_INIT; \
6831 sd->level = SD_LV_##type; \
6836 SD_INIT_FUNC(ALLNODES)
6839 #ifdef CONFIG_SCHED_SMT
6840 SD_INIT_FUNC(SIBLING)
6842 #ifdef CONFIG_SCHED_MC
6847 * To minimize stack usage kmalloc room for cpumasks and share the
6848 * space as the usage in build_sched_domains() dictates. Used only
6849 * if the amount of space is significant.
6852 cpumask_t tmpmask; /* make this one first */
6855 cpumask_t this_sibling_map;
6856 cpumask_t this_core_map;
6858 cpumask_t send_covered;
6861 cpumask_t domainspan;
6863 cpumask_t notcovered;
6868 #define SCHED_CPUMASK_ALLOC 1
6869 #define SCHED_CPUMASK_FREE(v) kfree(v)
6870 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
6872 #define SCHED_CPUMASK_ALLOC 0
6873 #define SCHED_CPUMASK_FREE(v)
6874 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
6877 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
6878 ((unsigned long)(a) + offsetof(struct allmasks, v))
6880 static int default_relax_domain_level = -1;
6882 static int __init setup_relax_domain_level(char *str)
6884 default_relax_domain_level = simple_strtoul(str, NULL, 0);
6887 __setup("relax_domain_level=", setup_relax_domain_level);
6889 static void set_domain_attribute(struct sched_domain *sd,
6890 struct sched_domain_attr *attr)
6894 if (!attr || attr->relax_domain_level < 0) {
6895 if (default_relax_domain_level < 0)
6898 request = default_relax_domain_level;
6900 request = attr->relax_domain_level;
6901 if (request < sd->level) {
6902 /* turn off idle balance on this domain */
6903 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
6905 /* turn on idle balance on this domain */
6906 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
6911 * Build sched domains for a given set of cpus and attach the sched domains
6912 * to the individual cpus
6914 static int __build_sched_domains(const cpumask_t *cpu_map,
6915 struct sched_domain_attr *attr)
6918 struct root_domain *rd;
6919 SCHED_CPUMASK_DECLARE(allmasks);
6922 struct sched_group **sched_group_nodes = NULL;
6923 int sd_allnodes = 0;
6926 * Allocate the per-node list of sched groups
6928 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6930 if (!sched_group_nodes) {
6931 printk(KERN_WARNING "Can not alloc sched group node list\n");
6936 rd = alloc_rootdomain();
6938 printk(KERN_WARNING "Cannot alloc root domain\n");
6940 kfree(sched_group_nodes);
6945 #if SCHED_CPUMASK_ALLOC
6946 /* get space for all scratch cpumask variables */
6947 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
6949 printk(KERN_WARNING "Cannot alloc cpumask array\n");
6952 kfree(sched_group_nodes);
6957 tmpmask = (cpumask_t *)allmasks;
6961 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
6965 * Set up domains for cpus specified by the cpu_map.
6967 for_each_cpu_mask(i, *cpu_map) {
6968 struct sched_domain *sd = NULL, *p;
6969 SCHED_CPUMASK_VAR(nodemask, allmasks);
6971 *nodemask = node_to_cpumask(cpu_to_node(i));
6972 cpus_and(*nodemask, *nodemask, *cpu_map);
6975 if (cpus_weight(*cpu_map) >
6976 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
6977 sd = &per_cpu(allnodes_domains, i);
6978 SD_INIT(sd, ALLNODES);
6979 set_domain_attribute(sd, attr);
6980 sd->span = *cpu_map;
6981 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
6987 sd = &per_cpu(node_domains, i);
6989 set_domain_attribute(sd, attr);
6990 sched_domain_node_span(cpu_to_node(i), &sd->span);
6994 cpus_and(sd->span, sd->span, *cpu_map);
6998 sd = &per_cpu(phys_domains, i);
7000 set_domain_attribute(sd, attr);
7001 sd->span = *nodemask;
7005 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7007 #ifdef CONFIG_SCHED_MC
7009 sd = &per_cpu(core_domains, i);
7011 set_domain_attribute(sd, attr);
7012 sd->span = cpu_coregroup_map(i);
7013 cpus_and(sd->span, sd->span, *cpu_map);
7016 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7019 #ifdef CONFIG_SCHED_SMT
7021 sd = &per_cpu(cpu_domains, i);
7022 SD_INIT(sd, SIBLING);
7023 set_domain_attribute(sd, attr);
7024 sd->span = per_cpu(cpu_sibling_map, i);
7025 cpus_and(sd->span, sd->span, *cpu_map);
7028 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7032 #ifdef CONFIG_SCHED_SMT
7033 /* Set up CPU (sibling) groups */
7034 for_each_cpu_mask(i, *cpu_map) {
7035 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7036 SCHED_CPUMASK_VAR(send_covered, allmasks);
7038 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7039 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7040 if (i != first_cpu(*this_sibling_map))
7043 init_sched_build_groups(this_sibling_map, cpu_map,
7045 send_covered, tmpmask);
7049 #ifdef CONFIG_SCHED_MC
7050 /* Set up multi-core groups */
7051 for_each_cpu_mask(i, *cpu_map) {
7052 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7053 SCHED_CPUMASK_VAR(send_covered, allmasks);
7055 *this_core_map = cpu_coregroup_map(i);
7056 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7057 if (i != first_cpu(*this_core_map))
7060 init_sched_build_groups(this_core_map, cpu_map,
7062 send_covered, tmpmask);
7066 /* Set up physical groups */
7067 for (i = 0; i < MAX_NUMNODES; i++) {
7068 SCHED_CPUMASK_VAR(nodemask, allmasks);
7069 SCHED_CPUMASK_VAR(send_covered, allmasks);
7071 *nodemask = node_to_cpumask(i);
7072 cpus_and(*nodemask, *nodemask, *cpu_map);
7073 if (cpus_empty(*nodemask))
7076 init_sched_build_groups(nodemask, cpu_map,
7078 send_covered, tmpmask);
7082 /* Set up node groups */
7084 SCHED_CPUMASK_VAR(send_covered, allmasks);
7086 init_sched_build_groups(cpu_map, cpu_map,
7087 &cpu_to_allnodes_group,
7088 send_covered, tmpmask);
7091 for (i = 0; i < MAX_NUMNODES; i++) {
7092 /* Set up node groups */
7093 struct sched_group *sg, *prev;
7094 SCHED_CPUMASK_VAR(nodemask, allmasks);
7095 SCHED_CPUMASK_VAR(domainspan, allmasks);
7096 SCHED_CPUMASK_VAR(covered, allmasks);
7099 *nodemask = node_to_cpumask(i);
7100 cpus_clear(*covered);
7102 cpus_and(*nodemask, *nodemask, *cpu_map);
7103 if (cpus_empty(*nodemask)) {
7104 sched_group_nodes[i] = NULL;
7108 sched_domain_node_span(i, domainspan);
7109 cpus_and(*domainspan, *domainspan, *cpu_map);
7111 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7113 printk(KERN_WARNING "Can not alloc domain group for "
7117 sched_group_nodes[i] = sg;
7118 for_each_cpu_mask(j, *nodemask) {
7119 struct sched_domain *sd;
7121 sd = &per_cpu(node_domains, j);
7124 sg->__cpu_power = 0;
7125 sg->cpumask = *nodemask;
7127 cpus_or(*covered, *covered, *nodemask);
7130 for (j = 0; j < MAX_NUMNODES; j++) {
7131 SCHED_CPUMASK_VAR(notcovered, allmasks);
7132 int n = (i + j) % MAX_NUMNODES;
7133 node_to_cpumask_ptr(pnodemask, n);
7135 cpus_complement(*notcovered, *covered);
7136 cpus_and(*tmpmask, *notcovered, *cpu_map);
7137 cpus_and(*tmpmask, *tmpmask, *domainspan);
7138 if (cpus_empty(*tmpmask))
7141 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7142 if (cpus_empty(*tmpmask))
7145 sg = kmalloc_node(sizeof(struct sched_group),
7149 "Can not alloc domain group for node %d\n", j);
7152 sg->__cpu_power = 0;
7153 sg->cpumask = *tmpmask;
7154 sg->next = prev->next;
7155 cpus_or(*covered, *covered, *tmpmask);
7162 /* Calculate CPU power for physical packages and nodes */
7163 #ifdef CONFIG_SCHED_SMT
7164 for_each_cpu_mask(i, *cpu_map) {
7165 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7167 init_sched_groups_power(i, sd);
7170 #ifdef CONFIG_SCHED_MC
7171 for_each_cpu_mask(i, *cpu_map) {
7172 struct sched_domain *sd = &per_cpu(core_domains, i);
7174 init_sched_groups_power(i, sd);
7178 for_each_cpu_mask(i, *cpu_map) {
7179 struct sched_domain *sd = &per_cpu(phys_domains, i);
7181 init_sched_groups_power(i, sd);
7185 for (i = 0; i < MAX_NUMNODES; i++)
7186 init_numa_sched_groups_power(sched_group_nodes[i]);
7189 struct sched_group *sg;
7191 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7193 init_numa_sched_groups_power(sg);
7197 /* Attach the domains */
7198 for_each_cpu_mask(i, *cpu_map) {
7199 struct sched_domain *sd;
7200 #ifdef CONFIG_SCHED_SMT
7201 sd = &per_cpu(cpu_domains, i);
7202 #elif defined(CONFIG_SCHED_MC)
7203 sd = &per_cpu(core_domains, i);
7205 sd = &per_cpu(phys_domains, i);
7207 cpu_attach_domain(sd, rd, i);
7210 SCHED_CPUMASK_FREE((void *)allmasks);
7215 free_sched_groups(cpu_map, tmpmask);
7216 SCHED_CPUMASK_FREE((void *)allmasks);
7221 static int build_sched_domains(const cpumask_t *cpu_map)
7223 return __build_sched_domains(cpu_map, NULL);
7226 static cpumask_t *doms_cur; /* current sched domains */
7227 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7228 static struct sched_domain_attr *dattr_cur;
7229 /* attribues of custom domains in 'doms_cur' */
7232 * Special case: If a kmalloc of a doms_cur partition (array of
7233 * cpumask_t) fails, then fallback to a single sched domain,
7234 * as determined by the single cpumask_t fallback_doms.
7236 static cpumask_t fallback_doms;
7238 void __attribute__((weak)) arch_update_cpu_topology(void)
7243 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7244 * For now this just excludes isolated cpus, but could be used to
7245 * exclude other special cases in the future.
7247 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7251 arch_update_cpu_topology();
7253 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7255 doms_cur = &fallback_doms;
7256 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7258 err = build_sched_domains(doms_cur);
7259 register_sched_domain_sysctl();
7264 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7267 free_sched_groups(cpu_map, tmpmask);
7271 * Detach sched domains from a group of cpus specified in cpu_map
7272 * These cpus will now be attached to the NULL domain
7274 static void detach_destroy_domains(const cpumask_t *cpu_map)
7279 unregister_sched_domain_sysctl();
7281 for_each_cpu_mask(i, *cpu_map)
7282 cpu_attach_domain(NULL, &def_root_domain, i);
7283 synchronize_sched();
7284 arch_destroy_sched_domains(cpu_map, &tmpmask);
7287 /* handle null as "default" */
7288 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7289 struct sched_domain_attr *new, int idx_new)
7291 struct sched_domain_attr tmp;
7298 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7299 new ? (new + idx_new) : &tmp,
7300 sizeof(struct sched_domain_attr));
7304 * Partition sched domains as specified by the 'ndoms_new'
7305 * cpumasks in the array doms_new[] of cpumasks. This compares
7306 * doms_new[] to the current sched domain partitioning, doms_cur[].
7307 * It destroys each deleted domain and builds each new domain.
7309 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7310 * The masks don't intersect (don't overlap.) We should setup one
7311 * sched domain for each mask. CPUs not in any of the cpumasks will
7312 * not be load balanced. If the same cpumask appears both in the
7313 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7316 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7317 * ownership of it and will kfree it when done with it. If the caller
7318 * failed the kmalloc call, then it can pass in doms_new == NULL,
7319 * and partition_sched_domains() will fallback to the single partition
7322 * Call with hotplug lock held
7324 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7325 struct sched_domain_attr *dattr_new)
7329 mutex_lock(&sched_domains_mutex);
7331 /* always unregister in case we don't destroy any domains */
7332 unregister_sched_domain_sysctl();
7334 if (doms_new == NULL) {
7336 doms_new = &fallback_doms;
7337 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7341 /* Destroy deleted domains */
7342 for (i = 0; i < ndoms_cur; i++) {
7343 for (j = 0; j < ndoms_new; j++) {
7344 if (cpus_equal(doms_cur[i], doms_new[j])
7345 && dattrs_equal(dattr_cur, i, dattr_new, j))
7348 /* no match - a current sched domain not in new doms_new[] */
7349 detach_destroy_domains(doms_cur + i);
7354 /* Build new domains */
7355 for (i = 0; i < ndoms_new; i++) {
7356 for (j = 0; j < ndoms_cur; j++) {
7357 if (cpus_equal(doms_new[i], doms_cur[j])
7358 && dattrs_equal(dattr_new, i, dattr_cur, j))
7361 /* no match - add a new doms_new */
7362 __build_sched_domains(doms_new + i,
7363 dattr_new ? dattr_new + i : NULL);
7368 /* Remember the new sched domains */
7369 if (doms_cur != &fallback_doms)
7371 kfree(dattr_cur); /* kfree(NULL) is safe */
7372 doms_cur = doms_new;
7373 dattr_cur = dattr_new;
7374 ndoms_cur = ndoms_new;
7376 register_sched_domain_sysctl();
7378 mutex_unlock(&sched_domains_mutex);
7381 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7382 int arch_reinit_sched_domains(void)
7387 mutex_lock(&sched_domains_mutex);
7388 detach_destroy_domains(&cpu_online_map);
7389 err = arch_init_sched_domains(&cpu_online_map);
7390 mutex_unlock(&sched_domains_mutex);
7396 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7400 if (buf[0] != '0' && buf[0] != '1')
7404 sched_smt_power_savings = (buf[0] == '1');
7406 sched_mc_power_savings = (buf[0] == '1');
7408 ret = arch_reinit_sched_domains();
7410 return ret ? ret : count;
7413 #ifdef CONFIG_SCHED_MC
7414 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7416 return sprintf(page, "%u\n", sched_mc_power_savings);
7418 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7419 const char *buf, size_t count)
7421 return sched_power_savings_store(buf, count, 0);
7423 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7424 sched_mc_power_savings_store);
7427 #ifdef CONFIG_SCHED_SMT
7428 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7430 return sprintf(page, "%u\n", sched_smt_power_savings);
7432 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7433 const char *buf, size_t count)
7435 return sched_power_savings_store(buf, count, 1);
7437 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7438 sched_smt_power_savings_store);
7441 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7445 #ifdef CONFIG_SCHED_SMT
7447 err = sysfs_create_file(&cls->kset.kobj,
7448 &attr_sched_smt_power_savings.attr);
7450 #ifdef CONFIG_SCHED_MC
7451 if (!err && mc_capable())
7452 err = sysfs_create_file(&cls->kset.kobj,
7453 &attr_sched_mc_power_savings.attr);
7460 * Force a reinitialization of the sched domains hierarchy. The domains
7461 * and groups cannot be updated in place without racing with the balancing
7462 * code, so we temporarily attach all running cpus to the NULL domain
7463 * which will prevent rebalancing while the sched domains are recalculated.
7465 static int update_sched_domains(struct notifier_block *nfb,
7466 unsigned long action, void *hcpu)
7469 case CPU_UP_PREPARE:
7470 case CPU_UP_PREPARE_FROZEN:
7471 case CPU_DOWN_PREPARE:
7472 case CPU_DOWN_PREPARE_FROZEN:
7473 detach_destroy_domains(&cpu_online_map);
7476 case CPU_UP_CANCELED:
7477 case CPU_UP_CANCELED_FROZEN:
7478 case CPU_DOWN_FAILED:
7479 case CPU_DOWN_FAILED_FROZEN:
7481 case CPU_ONLINE_FROZEN:
7483 case CPU_DEAD_FROZEN:
7485 * Fall through and re-initialise the domains.
7492 /* The hotplug lock is already held by cpu_up/cpu_down */
7493 arch_init_sched_domains(&cpu_online_map);
7498 void __init sched_init_smp(void)
7500 cpumask_t non_isolated_cpus;
7502 #if defined(CONFIG_NUMA)
7503 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7505 BUG_ON(sched_group_nodes_bycpu == NULL);
7508 mutex_lock(&sched_domains_mutex);
7509 arch_init_sched_domains(&cpu_online_map);
7510 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7511 if (cpus_empty(non_isolated_cpus))
7512 cpu_set(smp_processor_id(), non_isolated_cpus);
7513 mutex_unlock(&sched_domains_mutex);
7515 /* XXX: Theoretical race here - CPU may be hotplugged now */
7516 hotcpu_notifier(update_sched_domains, 0);
7519 /* Move init over to a non-isolated CPU */
7520 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7522 sched_init_granularity();
7525 void __init sched_init_smp(void)
7527 sched_init_granularity();
7529 #endif /* CONFIG_SMP */
7531 int in_sched_functions(unsigned long addr)
7533 return in_lock_functions(addr) ||
7534 (addr >= (unsigned long)__sched_text_start
7535 && addr < (unsigned long)__sched_text_end);
7538 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7540 cfs_rq->tasks_timeline = RB_ROOT;
7541 INIT_LIST_HEAD(&cfs_rq->tasks);
7542 #ifdef CONFIG_FAIR_GROUP_SCHED
7545 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7548 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7550 struct rt_prio_array *array;
7553 array = &rt_rq->active;
7554 for (i = 0; i < MAX_RT_PRIO; i++) {
7555 INIT_LIST_HEAD(array->xqueue + i);
7556 INIT_LIST_HEAD(array->squeue + i);
7557 __clear_bit(i, array->bitmap);
7559 /* delimiter for bitsearch: */
7560 __set_bit(MAX_RT_PRIO, array->bitmap);
7562 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7563 rt_rq->highest_prio = MAX_RT_PRIO;
7566 rt_rq->rt_nr_migratory = 0;
7567 rt_rq->overloaded = 0;
7571 rt_rq->rt_throttled = 0;
7572 rt_rq->rt_runtime = 0;
7573 spin_lock_init(&rt_rq->rt_runtime_lock);
7575 #ifdef CONFIG_RT_GROUP_SCHED
7576 rt_rq->rt_nr_boosted = 0;
7581 #ifdef CONFIG_FAIR_GROUP_SCHED
7582 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7583 struct sched_entity *se, int cpu, int add,
7584 struct sched_entity *parent)
7586 struct rq *rq = cpu_rq(cpu);
7587 tg->cfs_rq[cpu] = cfs_rq;
7588 init_cfs_rq(cfs_rq, rq);
7591 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7594 /* se could be NULL for init_task_group */
7599 se->cfs_rq = &rq->cfs;
7601 se->cfs_rq = parent->my_q;
7604 se->load.weight = tg->shares;
7605 se->load.inv_weight = 0;
7606 se->parent = parent;
7610 #ifdef CONFIG_RT_GROUP_SCHED
7611 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7612 struct sched_rt_entity *rt_se, int cpu, int add,
7613 struct sched_rt_entity *parent)
7615 struct rq *rq = cpu_rq(cpu);
7617 tg->rt_rq[cpu] = rt_rq;
7618 init_rt_rq(rt_rq, rq);
7620 rt_rq->rt_se = rt_se;
7621 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7623 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7625 tg->rt_se[cpu] = rt_se;
7630 rt_se->rt_rq = &rq->rt;
7632 rt_se->rt_rq = parent->my_q;
7634 rt_se->rt_rq = &rq->rt;
7635 rt_se->my_q = rt_rq;
7636 rt_se->parent = parent;
7637 INIT_LIST_HEAD(&rt_se->run_list);
7641 void __init sched_init(void)
7644 unsigned long alloc_size = 0, ptr;
7646 #ifdef CONFIG_FAIR_GROUP_SCHED
7647 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7649 #ifdef CONFIG_RT_GROUP_SCHED
7650 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7652 #ifdef CONFIG_USER_SCHED
7656 * As sched_init() is called before page_alloc is setup,
7657 * we use alloc_bootmem().
7660 ptr = (unsigned long)alloc_bootmem(alloc_size);
7662 #ifdef CONFIG_FAIR_GROUP_SCHED
7663 init_task_group.se = (struct sched_entity **)ptr;
7664 ptr += nr_cpu_ids * sizeof(void **);
7666 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7667 ptr += nr_cpu_ids * sizeof(void **);
7669 #ifdef CONFIG_USER_SCHED
7670 root_task_group.se = (struct sched_entity **)ptr;
7671 ptr += nr_cpu_ids * sizeof(void **);
7673 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7674 ptr += nr_cpu_ids * sizeof(void **);
7677 #ifdef CONFIG_RT_GROUP_SCHED
7678 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7679 ptr += nr_cpu_ids * sizeof(void **);
7681 init_task_group.rt_rq = (struct rt_rq **)ptr;
7682 ptr += nr_cpu_ids * sizeof(void **);
7684 #ifdef CONFIG_USER_SCHED
7685 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7686 ptr += nr_cpu_ids * sizeof(void **);
7688 root_task_group.rt_rq = (struct rt_rq **)ptr;
7689 ptr += nr_cpu_ids * sizeof(void **);
7695 init_defrootdomain();
7698 init_rt_bandwidth(&def_rt_bandwidth,
7699 global_rt_period(), global_rt_runtime());
7701 #ifdef CONFIG_RT_GROUP_SCHED
7702 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7703 global_rt_period(), global_rt_runtime());
7704 #ifdef CONFIG_USER_SCHED
7705 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7706 global_rt_period(), RUNTIME_INF);
7710 #ifdef CONFIG_GROUP_SCHED
7711 list_add(&init_task_group.list, &task_groups);
7712 INIT_LIST_HEAD(&init_task_group.children);
7714 #ifdef CONFIG_USER_SCHED
7715 INIT_LIST_HEAD(&root_task_group.children);
7716 init_task_group.parent = &root_task_group;
7717 list_add(&init_task_group.siblings, &root_task_group.children);
7721 for_each_possible_cpu(i) {
7725 spin_lock_init(&rq->lock);
7726 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7728 init_cfs_rq(&rq->cfs, rq);
7729 init_rt_rq(&rq->rt, rq);
7730 #ifdef CONFIG_FAIR_GROUP_SCHED
7731 init_task_group.shares = init_task_group_load;
7732 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7733 #ifdef CONFIG_CGROUP_SCHED
7735 * How much cpu bandwidth does init_task_group get?
7737 * In case of task-groups formed thr' the cgroup filesystem, it
7738 * gets 100% of the cpu resources in the system. This overall
7739 * system cpu resource is divided among the tasks of
7740 * init_task_group and its child task-groups in a fair manner,
7741 * based on each entity's (task or task-group's) weight
7742 * (se->load.weight).
7744 * In other words, if init_task_group has 10 tasks of weight
7745 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7746 * then A0's share of the cpu resource is:
7748 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7750 * We achieve this by letting init_task_group's tasks sit
7751 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7753 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7754 #elif defined CONFIG_USER_SCHED
7755 root_task_group.shares = NICE_0_LOAD;
7756 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
7758 * In case of task-groups formed thr' the user id of tasks,
7759 * init_task_group represents tasks belonging to root user.
7760 * Hence it forms a sibling of all subsequent groups formed.
7761 * In this case, init_task_group gets only a fraction of overall
7762 * system cpu resource, based on the weight assigned to root
7763 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7764 * by letting tasks of init_task_group sit in a separate cfs_rq
7765 * (init_cfs_rq) and having one entity represent this group of
7766 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7768 init_tg_cfs_entry(&init_task_group,
7769 &per_cpu(init_cfs_rq, i),
7770 &per_cpu(init_sched_entity, i), i, 1,
7771 root_task_group.se[i]);
7774 #endif /* CONFIG_FAIR_GROUP_SCHED */
7776 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7777 #ifdef CONFIG_RT_GROUP_SCHED
7778 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7779 #ifdef CONFIG_CGROUP_SCHED
7780 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7781 #elif defined CONFIG_USER_SCHED
7782 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
7783 init_tg_rt_entry(&init_task_group,
7784 &per_cpu(init_rt_rq, i),
7785 &per_cpu(init_sched_rt_entity, i), i, 1,
7786 root_task_group.rt_se[i]);
7790 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7791 rq->cpu_load[j] = 0;
7795 rq->active_balance = 0;
7796 rq->next_balance = jiffies;
7799 rq->migration_thread = NULL;
7800 INIT_LIST_HEAD(&rq->migration_queue);
7801 rq_attach_root(rq, &def_root_domain);
7804 atomic_set(&rq->nr_iowait, 0);
7807 set_load_weight(&init_task);
7809 #ifdef CONFIG_PREEMPT_NOTIFIERS
7810 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7814 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7817 #ifdef CONFIG_RT_MUTEXES
7818 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7822 * The boot idle thread does lazy MMU switching as well:
7824 atomic_inc(&init_mm.mm_count);
7825 enter_lazy_tlb(&init_mm, current);
7828 * Make us the idle thread. Technically, schedule() should not be
7829 * called from this thread, however somewhere below it might be,
7830 * but because we are the idle thread, we just pick up running again
7831 * when this runqueue becomes "idle".
7833 init_idle(current, smp_processor_id());
7835 * During early bootup we pretend to be a normal task:
7837 current->sched_class = &fair_sched_class;
7839 scheduler_running = 1;
7842 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7843 void __might_sleep(char *file, int line)
7846 static unsigned long prev_jiffy; /* ratelimiting */
7848 if ((in_atomic() || irqs_disabled()) &&
7849 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7850 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7852 prev_jiffy = jiffies;
7853 printk(KERN_ERR "BUG: sleeping function called from invalid"
7854 " context at %s:%d\n", file, line);
7855 printk("in_atomic():%d, irqs_disabled():%d\n",
7856 in_atomic(), irqs_disabled());
7857 debug_show_held_locks(current);
7858 if (irqs_disabled())
7859 print_irqtrace_events(current);
7864 EXPORT_SYMBOL(__might_sleep);
7867 #ifdef CONFIG_MAGIC_SYSRQ
7868 static void normalize_task(struct rq *rq, struct task_struct *p)
7872 update_rq_clock(rq);
7873 on_rq = p->se.on_rq;
7875 deactivate_task(rq, p, 0);
7876 __setscheduler(rq, p, SCHED_NORMAL, 0);
7878 activate_task(rq, p, 0);
7879 resched_task(rq->curr);
7883 void normalize_rt_tasks(void)
7885 struct task_struct *g, *p;
7886 unsigned long flags;
7889 read_lock_irqsave(&tasklist_lock, flags);
7890 do_each_thread(g, p) {
7892 * Only normalize user tasks:
7897 p->se.exec_start = 0;
7898 #ifdef CONFIG_SCHEDSTATS
7899 p->se.wait_start = 0;
7900 p->se.sleep_start = 0;
7901 p->se.block_start = 0;
7906 * Renice negative nice level userspace
7909 if (TASK_NICE(p) < 0 && p->mm)
7910 set_user_nice(p, 0);
7914 spin_lock(&p->pi_lock);
7915 rq = __task_rq_lock(p);
7917 normalize_task(rq, p);
7919 __task_rq_unlock(rq);
7920 spin_unlock(&p->pi_lock);
7921 } while_each_thread(g, p);
7923 read_unlock_irqrestore(&tasklist_lock, flags);
7926 #endif /* CONFIG_MAGIC_SYSRQ */
7930 * These functions are only useful for the IA64 MCA handling.
7932 * They can only be called when the whole system has been
7933 * stopped - every CPU needs to be quiescent, and no scheduling
7934 * activity can take place. Using them for anything else would
7935 * be a serious bug, and as a result, they aren't even visible
7936 * under any other configuration.
7940 * curr_task - return the current task for a given cpu.
7941 * @cpu: the processor in question.
7943 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7945 struct task_struct *curr_task(int cpu)
7947 return cpu_curr(cpu);
7951 * set_curr_task - set the current task for a given cpu.
7952 * @cpu: the processor in question.
7953 * @p: the task pointer to set.
7955 * Description: This function must only be used when non-maskable interrupts
7956 * are serviced on a separate stack. It allows the architecture to switch the
7957 * notion of the current task on a cpu in a non-blocking manner. This function
7958 * must be called with all CPU's synchronized, and interrupts disabled, the
7959 * and caller must save the original value of the current task (see
7960 * curr_task() above) and restore that value before reenabling interrupts and
7961 * re-starting the system.
7963 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7965 void set_curr_task(int cpu, struct task_struct *p)
7972 #ifdef CONFIG_FAIR_GROUP_SCHED
7973 static void free_fair_sched_group(struct task_group *tg)
7977 for_each_possible_cpu(i) {
7979 kfree(tg->cfs_rq[i]);
7989 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7991 struct cfs_rq *cfs_rq;
7992 struct sched_entity *se, *parent_se;
7996 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7999 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8003 tg->shares = NICE_0_LOAD;
8005 for_each_possible_cpu(i) {
8008 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8009 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8013 se = kmalloc_node(sizeof(struct sched_entity),
8014 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8018 parent_se = parent ? parent->se[i] : NULL;
8019 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8028 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8030 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8031 &cpu_rq(cpu)->leaf_cfs_rq_list);
8034 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8036 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8039 static inline void free_fair_sched_group(struct task_group *tg)
8044 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8049 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8053 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8058 #ifdef CONFIG_RT_GROUP_SCHED
8059 static void free_rt_sched_group(struct task_group *tg)
8063 destroy_rt_bandwidth(&tg->rt_bandwidth);
8065 for_each_possible_cpu(i) {
8067 kfree(tg->rt_rq[i]);
8069 kfree(tg->rt_se[i]);
8077 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8079 struct rt_rq *rt_rq;
8080 struct sched_rt_entity *rt_se, *parent_se;
8084 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8087 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8091 init_rt_bandwidth(&tg->rt_bandwidth,
8092 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8094 for_each_possible_cpu(i) {
8097 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8098 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8102 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8103 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8107 parent_se = parent ? parent->rt_se[i] : NULL;
8108 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8117 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8119 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8120 &cpu_rq(cpu)->leaf_rt_rq_list);
8123 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8125 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8128 static inline void free_rt_sched_group(struct task_group *tg)
8133 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8138 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8142 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8147 #ifdef CONFIG_GROUP_SCHED
8148 static void free_sched_group(struct task_group *tg)
8150 free_fair_sched_group(tg);
8151 free_rt_sched_group(tg);
8155 /* allocate runqueue etc for a new task group */
8156 struct task_group *sched_create_group(struct task_group *parent)
8158 struct task_group *tg;
8159 unsigned long flags;
8162 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8164 return ERR_PTR(-ENOMEM);
8166 if (!alloc_fair_sched_group(tg, parent))
8169 if (!alloc_rt_sched_group(tg, parent))
8172 spin_lock_irqsave(&task_group_lock, flags);
8173 for_each_possible_cpu(i) {
8174 register_fair_sched_group(tg, i);
8175 register_rt_sched_group(tg, i);
8177 list_add_rcu(&tg->list, &task_groups);
8179 WARN_ON(!parent); /* root should already exist */
8181 tg->parent = parent;
8182 list_add_rcu(&tg->siblings, &parent->children);
8183 INIT_LIST_HEAD(&tg->children);
8184 spin_unlock_irqrestore(&task_group_lock, flags);
8189 free_sched_group(tg);
8190 return ERR_PTR(-ENOMEM);
8193 /* rcu callback to free various structures associated with a task group */
8194 static void free_sched_group_rcu(struct rcu_head *rhp)
8196 /* now it should be safe to free those cfs_rqs */
8197 free_sched_group(container_of(rhp, struct task_group, rcu));
8200 /* Destroy runqueue etc associated with a task group */
8201 void sched_destroy_group(struct task_group *tg)
8203 unsigned long flags;
8206 spin_lock_irqsave(&task_group_lock, flags);
8207 for_each_possible_cpu(i) {
8208 unregister_fair_sched_group(tg, i);
8209 unregister_rt_sched_group(tg, i);
8211 list_del_rcu(&tg->list);
8212 list_del_rcu(&tg->siblings);
8213 spin_unlock_irqrestore(&task_group_lock, flags);
8215 /* wait for possible concurrent references to cfs_rqs complete */
8216 call_rcu(&tg->rcu, free_sched_group_rcu);
8219 /* change task's runqueue when it moves between groups.
8220 * The caller of this function should have put the task in its new group
8221 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8222 * reflect its new group.
8224 void sched_move_task(struct task_struct *tsk)
8227 unsigned long flags;
8230 rq = task_rq_lock(tsk, &flags);
8232 update_rq_clock(rq);
8234 running = task_current(rq, tsk);
8235 on_rq = tsk->se.on_rq;
8238 dequeue_task(rq, tsk, 0);
8239 if (unlikely(running))
8240 tsk->sched_class->put_prev_task(rq, tsk);
8242 set_task_rq(tsk, task_cpu(tsk));
8244 #ifdef CONFIG_FAIR_GROUP_SCHED
8245 if (tsk->sched_class->moved_group)
8246 tsk->sched_class->moved_group(tsk);
8249 if (unlikely(running))
8250 tsk->sched_class->set_curr_task(rq);
8252 enqueue_task(rq, tsk, 0);
8254 task_rq_unlock(rq, &flags);
8258 #ifdef CONFIG_FAIR_GROUP_SCHED
8259 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8261 struct cfs_rq *cfs_rq = se->cfs_rq;
8262 struct rq *rq = cfs_rq->rq;
8265 spin_lock_irq(&rq->lock);
8269 dequeue_entity(cfs_rq, se, 0);
8271 se->load.weight = shares;
8272 se->load.inv_weight = 0;
8275 enqueue_entity(cfs_rq, se, 0);
8277 spin_unlock_irq(&rq->lock);
8280 static DEFINE_MUTEX(shares_mutex);
8282 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8285 unsigned long flags;
8288 * We can't change the weight of the root cgroup.
8293 if (shares < MIN_SHARES)
8294 shares = MIN_SHARES;
8295 else if (shares > MAX_SHARES)
8296 shares = MAX_SHARES;
8298 mutex_lock(&shares_mutex);
8299 if (tg->shares == shares)
8302 spin_lock_irqsave(&task_group_lock, flags);
8303 for_each_possible_cpu(i)
8304 unregister_fair_sched_group(tg, i);
8305 list_del_rcu(&tg->siblings);
8306 spin_unlock_irqrestore(&task_group_lock, flags);
8308 /* wait for any ongoing reference to this group to finish */
8309 synchronize_sched();
8312 * Now we are free to modify the group's share on each cpu
8313 * w/o tripping rebalance_share or load_balance_fair.
8315 tg->shares = shares;
8316 for_each_possible_cpu(i)
8317 set_se_shares(tg->se[i], shares);
8320 * Enable load balance activity on this group, by inserting it back on
8321 * each cpu's rq->leaf_cfs_rq_list.
8323 spin_lock_irqsave(&task_group_lock, flags);
8324 for_each_possible_cpu(i)
8325 register_fair_sched_group(tg, i);
8326 list_add_rcu(&tg->siblings, &tg->parent->children);
8327 spin_unlock_irqrestore(&task_group_lock, flags);
8329 mutex_unlock(&shares_mutex);
8333 unsigned long sched_group_shares(struct task_group *tg)
8339 #ifdef CONFIG_RT_GROUP_SCHED
8341 * Ensure that the real time constraints are schedulable.
8343 static DEFINE_MUTEX(rt_constraints_mutex);
8345 static unsigned long to_ratio(u64 period, u64 runtime)
8347 if (runtime == RUNTIME_INF)
8350 return div64_u64(runtime << 16, period);
8353 #ifdef CONFIG_CGROUP_SCHED
8354 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8356 struct task_group *tgi, *parent = tg->parent;
8357 unsigned long total = 0;
8360 if (global_rt_period() < period)
8363 return to_ratio(period, runtime) <
8364 to_ratio(global_rt_period(), global_rt_runtime());
8367 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8371 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8375 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8376 tgi->rt_bandwidth.rt_runtime);
8380 return total + to_ratio(period, runtime) <
8381 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8382 parent->rt_bandwidth.rt_runtime);
8384 #elif defined CONFIG_USER_SCHED
8385 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8387 struct task_group *tgi;
8388 unsigned long total = 0;
8389 unsigned long global_ratio =
8390 to_ratio(global_rt_period(), global_rt_runtime());
8393 list_for_each_entry_rcu(tgi, &task_groups, list) {
8397 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8398 tgi->rt_bandwidth.rt_runtime);
8402 return total + to_ratio(period, runtime) < global_ratio;
8406 /* Must be called with tasklist_lock held */
8407 static inline int tg_has_rt_tasks(struct task_group *tg)
8409 struct task_struct *g, *p;
8410 do_each_thread(g, p) {
8411 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8413 } while_each_thread(g, p);
8417 static int tg_set_bandwidth(struct task_group *tg,
8418 u64 rt_period, u64 rt_runtime)
8422 mutex_lock(&rt_constraints_mutex);
8423 read_lock(&tasklist_lock);
8424 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8428 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8433 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8434 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8435 tg->rt_bandwidth.rt_runtime = rt_runtime;
8437 for_each_possible_cpu(i) {
8438 struct rt_rq *rt_rq = tg->rt_rq[i];
8440 spin_lock(&rt_rq->rt_runtime_lock);
8441 rt_rq->rt_runtime = rt_runtime;
8442 spin_unlock(&rt_rq->rt_runtime_lock);
8444 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8446 read_unlock(&tasklist_lock);
8447 mutex_unlock(&rt_constraints_mutex);
8452 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8454 u64 rt_runtime, rt_period;
8456 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8457 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8458 if (rt_runtime_us < 0)
8459 rt_runtime = RUNTIME_INF;
8461 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8464 long sched_group_rt_runtime(struct task_group *tg)
8468 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8471 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8472 do_div(rt_runtime_us, NSEC_PER_USEC);
8473 return rt_runtime_us;
8476 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8478 u64 rt_runtime, rt_period;
8480 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8481 rt_runtime = tg->rt_bandwidth.rt_runtime;
8483 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8486 long sched_group_rt_period(struct task_group *tg)
8490 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8491 do_div(rt_period_us, NSEC_PER_USEC);
8492 return rt_period_us;
8495 static int sched_rt_global_constraints(void)
8499 mutex_lock(&rt_constraints_mutex);
8500 if (!__rt_schedulable(NULL, 1, 0))
8502 mutex_unlock(&rt_constraints_mutex);
8507 static int sched_rt_global_constraints(void)
8509 unsigned long flags;
8512 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8513 for_each_possible_cpu(i) {
8514 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8516 spin_lock(&rt_rq->rt_runtime_lock);
8517 rt_rq->rt_runtime = global_rt_runtime();
8518 spin_unlock(&rt_rq->rt_runtime_lock);
8520 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8526 int sched_rt_handler(struct ctl_table *table, int write,
8527 struct file *filp, void __user *buffer, size_t *lenp,
8531 int old_period, old_runtime;
8532 static DEFINE_MUTEX(mutex);
8535 old_period = sysctl_sched_rt_period;
8536 old_runtime = sysctl_sched_rt_runtime;
8538 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8540 if (!ret && write) {
8541 ret = sched_rt_global_constraints();
8543 sysctl_sched_rt_period = old_period;
8544 sysctl_sched_rt_runtime = old_runtime;
8546 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8547 def_rt_bandwidth.rt_period =
8548 ns_to_ktime(global_rt_period());
8551 mutex_unlock(&mutex);
8556 #ifdef CONFIG_CGROUP_SCHED
8558 /* return corresponding task_group object of a cgroup */
8559 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8561 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8562 struct task_group, css);
8565 static struct cgroup_subsys_state *
8566 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8568 struct task_group *tg, *parent;
8570 if (!cgrp->parent) {
8571 /* This is early initialization for the top cgroup */
8572 init_task_group.css.cgroup = cgrp;
8573 return &init_task_group.css;
8576 parent = cgroup_tg(cgrp->parent);
8577 tg = sched_create_group(parent);
8579 return ERR_PTR(-ENOMEM);
8581 /* Bind the cgroup to task_group object we just created */
8582 tg->css.cgroup = cgrp;
8588 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8590 struct task_group *tg = cgroup_tg(cgrp);
8592 sched_destroy_group(tg);
8596 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8597 struct task_struct *tsk)
8599 #ifdef CONFIG_RT_GROUP_SCHED
8600 /* Don't accept realtime tasks when there is no way for them to run */
8601 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8604 /* We don't support RT-tasks being in separate groups */
8605 if (tsk->sched_class != &fair_sched_class)
8613 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8614 struct cgroup *old_cont, struct task_struct *tsk)
8616 sched_move_task(tsk);
8619 #ifdef CONFIG_FAIR_GROUP_SCHED
8620 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8623 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8626 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8628 struct task_group *tg = cgroup_tg(cgrp);
8630 return (u64) tg->shares;
8634 #ifdef CONFIG_RT_GROUP_SCHED
8635 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8638 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8641 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8643 return sched_group_rt_runtime(cgroup_tg(cgrp));
8646 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8649 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8652 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8654 return sched_group_rt_period(cgroup_tg(cgrp));
8658 static struct cftype cpu_files[] = {
8659 #ifdef CONFIG_FAIR_GROUP_SCHED
8662 .read_u64 = cpu_shares_read_u64,
8663 .write_u64 = cpu_shares_write_u64,
8666 #ifdef CONFIG_RT_GROUP_SCHED
8668 .name = "rt_runtime_us",
8669 .read_s64 = cpu_rt_runtime_read,
8670 .write_s64 = cpu_rt_runtime_write,
8673 .name = "rt_period_us",
8674 .read_u64 = cpu_rt_period_read_uint,
8675 .write_u64 = cpu_rt_period_write_uint,
8680 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8682 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8685 struct cgroup_subsys cpu_cgroup_subsys = {
8687 .create = cpu_cgroup_create,
8688 .destroy = cpu_cgroup_destroy,
8689 .can_attach = cpu_cgroup_can_attach,
8690 .attach = cpu_cgroup_attach,
8691 .populate = cpu_cgroup_populate,
8692 .subsys_id = cpu_cgroup_subsys_id,
8696 #endif /* CONFIG_CGROUP_SCHED */
8698 #ifdef CONFIG_CGROUP_CPUACCT
8701 * CPU accounting code for task groups.
8703 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8704 * (balbir@in.ibm.com).
8707 /* track cpu usage of a group of tasks */
8709 struct cgroup_subsys_state css;
8710 /* cpuusage holds pointer to a u64-type object on every cpu */
8714 struct cgroup_subsys cpuacct_subsys;
8716 /* return cpu accounting group corresponding to this container */
8717 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8719 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8720 struct cpuacct, css);
8723 /* return cpu accounting group to which this task belongs */
8724 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8726 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8727 struct cpuacct, css);
8730 /* create a new cpu accounting group */
8731 static struct cgroup_subsys_state *cpuacct_create(
8732 struct cgroup_subsys *ss, struct cgroup *cgrp)
8734 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8737 return ERR_PTR(-ENOMEM);
8739 ca->cpuusage = alloc_percpu(u64);
8740 if (!ca->cpuusage) {
8742 return ERR_PTR(-ENOMEM);
8748 /* destroy an existing cpu accounting group */
8750 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8752 struct cpuacct *ca = cgroup_ca(cgrp);
8754 free_percpu(ca->cpuusage);
8758 /* return total cpu usage (in nanoseconds) of a group */
8759 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8761 struct cpuacct *ca = cgroup_ca(cgrp);
8762 u64 totalcpuusage = 0;
8765 for_each_possible_cpu(i) {
8766 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8769 * Take rq->lock to make 64-bit addition safe on 32-bit
8772 spin_lock_irq(&cpu_rq(i)->lock);
8773 totalcpuusage += *cpuusage;
8774 spin_unlock_irq(&cpu_rq(i)->lock);
8777 return totalcpuusage;
8780 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8783 struct cpuacct *ca = cgroup_ca(cgrp);
8792 for_each_possible_cpu(i) {
8793 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8795 spin_lock_irq(&cpu_rq(i)->lock);
8797 spin_unlock_irq(&cpu_rq(i)->lock);
8803 static struct cftype files[] = {
8806 .read_u64 = cpuusage_read,
8807 .write_u64 = cpuusage_write,
8811 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8813 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8817 * charge this task's execution time to its accounting group.
8819 * called with rq->lock held.
8821 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8825 if (!cpuacct_subsys.active)
8830 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8832 *cpuusage += cputime;
8836 struct cgroup_subsys cpuacct_subsys = {
8838 .create = cpuacct_create,
8839 .destroy = cpuacct_destroy,
8840 .populate = cpuacct_populate,
8841 .subsys_id = cpuacct_subsys_id,
8843 #endif /* CONFIG_CGROUP_CPUACCT */