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 queue[MAX_RT_PRIO];
159 struct rt_bandwidth {
160 /* nests inside the rq lock: */
161 spinlock_t rt_runtime_lock;
164 struct hrtimer rt_period_timer;
167 static struct rt_bandwidth def_rt_bandwidth;
169 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
171 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
173 struct rt_bandwidth *rt_b =
174 container_of(timer, struct rt_bandwidth, rt_period_timer);
180 now = hrtimer_cb_get_time(timer);
181 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
186 idle = do_sched_rt_period_timer(rt_b, overrun);
189 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
193 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
195 rt_b->rt_period = ns_to_ktime(period);
196 rt_b->rt_runtime = runtime;
198 spin_lock_init(&rt_b->rt_runtime_lock);
200 hrtimer_init(&rt_b->rt_period_timer,
201 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
202 rt_b->rt_period_timer.function = sched_rt_period_timer;
203 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
206 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
210 if (rt_b->rt_runtime == RUNTIME_INF)
213 if (hrtimer_active(&rt_b->rt_period_timer))
216 spin_lock(&rt_b->rt_runtime_lock);
218 if (hrtimer_active(&rt_b->rt_period_timer))
221 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
222 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
223 hrtimer_start(&rt_b->rt_period_timer,
224 rt_b->rt_period_timer.expires,
227 spin_unlock(&rt_b->rt_runtime_lock);
230 #ifdef CONFIG_RT_GROUP_SCHED
231 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
233 hrtimer_cancel(&rt_b->rt_period_timer);
238 * sched_domains_mutex serializes calls to arch_init_sched_domains,
239 * detach_destroy_domains and partition_sched_domains.
241 static DEFINE_MUTEX(sched_domains_mutex);
243 #ifdef CONFIG_GROUP_SCHED
245 #include <linux/cgroup.h>
249 static LIST_HEAD(task_groups);
251 /* task group related information */
253 #ifdef CONFIG_CGROUP_SCHED
254 struct cgroup_subsys_state css;
257 #ifdef CONFIG_FAIR_GROUP_SCHED
258 /* schedulable entities of this group on each cpu */
259 struct sched_entity **se;
260 /* runqueue "owned" by this group on each cpu */
261 struct cfs_rq **cfs_rq;
262 unsigned long shares;
265 #ifdef CONFIG_RT_GROUP_SCHED
266 struct sched_rt_entity **rt_se;
267 struct rt_rq **rt_rq;
269 struct rt_bandwidth rt_bandwidth;
273 struct list_head list;
275 struct task_group *parent;
276 struct list_head siblings;
277 struct list_head children;
280 #ifdef CONFIG_USER_SCHED
284 * Every UID task group (including init_task_group aka UID-0) will
285 * be a child to this group.
287 struct task_group root_task_group;
289 #ifdef CONFIG_FAIR_GROUP_SCHED
290 /* Default task group's sched entity on each cpu */
291 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
292 /* Default task group's cfs_rq on each cpu */
293 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
294 #endif /* CONFIG_FAIR_GROUP_SCHED */
296 #ifdef CONFIG_RT_GROUP_SCHED
297 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
298 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
299 #endif /* CONFIG_RT_GROUP_SCHED */
300 #else /* !CONFIG_FAIR_GROUP_SCHED */
301 #define root_task_group init_task_group
302 #endif /* CONFIG_FAIR_GROUP_SCHED */
304 /* task_group_lock serializes add/remove of task groups and also changes to
305 * a task group's cpu shares.
307 static DEFINE_SPINLOCK(task_group_lock);
309 #ifdef CONFIG_FAIR_GROUP_SCHED
310 #ifdef CONFIG_USER_SCHED
311 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
312 #else /* !CONFIG_USER_SCHED */
313 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
314 #endif /* CONFIG_USER_SCHED */
317 * A weight of 0 or 1 can cause arithmetics problems.
318 * A weight of a cfs_rq is the sum of weights of which entities
319 * are queued on this cfs_rq, so a weight of a entity should not be
320 * too large, so as the shares value of a task group.
321 * (The default weight is 1024 - so there's no practical
322 * limitation from this.)
325 #define MAX_SHARES (1UL << 18)
327 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
330 /* Default task group.
331 * Every task in system belong to this group at bootup.
333 struct task_group init_task_group;
335 /* return group to which a task belongs */
336 static inline struct task_group *task_group(struct task_struct *p)
338 struct task_group *tg;
340 #ifdef CONFIG_USER_SCHED
342 #elif defined(CONFIG_CGROUP_SCHED)
343 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
344 struct task_group, css);
346 tg = &init_task_group;
351 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
352 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
354 #ifdef CONFIG_FAIR_GROUP_SCHED
355 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
356 p->se.parent = task_group(p)->se[cpu];
359 #ifdef CONFIG_RT_GROUP_SCHED
360 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
361 p->rt.parent = task_group(p)->rt_se[cpu];
367 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
369 #endif /* CONFIG_GROUP_SCHED */
371 /* CFS-related fields in a runqueue */
373 struct load_weight load;
374 unsigned long nr_running;
379 struct rb_root tasks_timeline;
380 struct rb_node *rb_leftmost;
382 struct list_head tasks;
383 struct list_head *balance_iterator;
386 * 'curr' points to currently running entity on this cfs_rq.
387 * It is set to NULL otherwise (i.e when none are currently running).
389 struct sched_entity *curr, *next;
391 unsigned long nr_spread_over;
393 #ifdef CONFIG_FAIR_GROUP_SCHED
394 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
397 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
398 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
399 * (like users, containers etc.)
401 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
402 * list is used during load balance.
404 struct list_head leaf_cfs_rq_list;
405 struct task_group *tg; /* group that "owns" this runqueue */
409 /* Real-Time classes' related field in a runqueue: */
411 struct rt_prio_array active;
412 unsigned long rt_nr_running;
413 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
414 int highest_prio; /* highest queued rt task prio */
417 unsigned long rt_nr_migratory;
423 /* Nests inside the rq lock: */
424 spinlock_t rt_runtime_lock;
426 #ifdef CONFIG_RT_GROUP_SCHED
427 unsigned long rt_nr_boosted;
430 struct list_head leaf_rt_rq_list;
431 struct task_group *tg;
432 struct sched_rt_entity *rt_se;
439 * We add the notion of a root-domain which will be used to define per-domain
440 * variables. Each exclusive cpuset essentially defines an island domain by
441 * fully partitioning the member cpus from any other cpuset. Whenever a new
442 * exclusive cpuset is created, we also create and attach a new root-domain
452 * The "RT overload" flag: it gets set if a CPU has more than
453 * one runnable RT task.
458 struct cpupri cpupri;
463 * By default the system creates a single root-domain with all cpus as
464 * members (mimicking the global state we have today).
466 static struct root_domain def_root_domain;
471 * This is the main, per-CPU runqueue data structure.
473 * Locking rule: those places that want to lock multiple runqueues
474 * (such as the load balancing or the thread migration code), lock
475 * acquire operations must be ordered by ascending &runqueue.
482 * nr_running and cpu_load should be in the same cacheline because
483 * remote CPUs use both these fields when doing load calculation.
485 unsigned long nr_running;
486 #define CPU_LOAD_IDX_MAX 5
487 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
488 unsigned char idle_at_tick;
490 unsigned long last_tick_seen;
491 unsigned char in_nohz_recently;
493 /* capture load from *all* tasks on this cpu: */
494 struct load_weight load;
495 unsigned long nr_load_updates;
501 #ifdef CONFIG_FAIR_GROUP_SCHED
502 /* list of leaf cfs_rq on this cpu: */
503 struct list_head leaf_cfs_rq_list;
505 #ifdef CONFIG_RT_GROUP_SCHED
506 struct list_head leaf_rt_rq_list;
510 * This is part of a global counter where only the total sum
511 * over all CPUs matters. A task can increase this counter on
512 * one CPU and if it got migrated afterwards it may decrease
513 * it on another CPU. Always updated under the runqueue lock:
515 unsigned long nr_uninterruptible;
517 struct task_struct *curr, *idle;
518 unsigned long next_balance;
519 struct mm_struct *prev_mm;
526 struct root_domain *rd;
527 struct sched_domain *sd;
529 /* For active balancing */
532 /* cpu of this runqueue: */
536 struct task_struct *migration_thread;
537 struct list_head migration_queue;
540 #ifdef CONFIG_SCHED_HRTICK
541 unsigned long hrtick_flags;
542 ktime_t hrtick_expire;
543 struct hrtimer hrtick_timer;
546 #ifdef CONFIG_SCHEDSTATS
548 struct sched_info rq_sched_info;
550 /* sys_sched_yield() stats */
551 unsigned int yld_exp_empty;
552 unsigned int yld_act_empty;
553 unsigned int yld_both_empty;
554 unsigned int yld_count;
556 /* schedule() stats */
557 unsigned int sched_switch;
558 unsigned int sched_count;
559 unsigned int sched_goidle;
561 /* try_to_wake_up() stats */
562 unsigned int ttwu_count;
563 unsigned int ttwu_local;
566 unsigned int bkl_count;
568 struct lock_class_key rq_lock_key;
571 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
573 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
575 rq->curr->sched_class->check_preempt_curr(rq, p);
578 static inline int cpu_of(struct rq *rq)
588 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
589 * See detach_destroy_domains: synchronize_sched for details.
591 * The domain tree of any CPU may only be accessed from within
592 * preempt-disabled sections.
594 #define for_each_domain(cpu, __sd) \
595 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
597 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
598 #define this_rq() (&__get_cpu_var(runqueues))
599 #define task_rq(p) cpu_rq(task_cpu(p))
600 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
602 static inline void update_rq_clock(struct rq *rq)
604 rq->clock = sched_clock_cpu(cpu_of(rq));
608 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
610 #ifdef CONFIG_SCHED_DEBUG
611 # define const_debug __read_mostly
613 # define const_debug static const
617 * Debugging: various feature bits
620 #define SCHED_FEAT(name, enabled) \
621 __SCHED_FEAT_##name ,
624 #include "sched_features.h"
629 #define SCHED_FEAT(name, enabled) \
630 (1UL << __SCHED_FEAT_##name) * enabled |
632 const_debug unsigned int sysctl_sched_features =
633 #include "sched_features.h"
638 #ifdef CONFIG_SCHED_DEBUG
639 #define SCHED_FEAT(name, enabled) \
642 static __read_mostly char *sched_feat_names[] = {
643 #include "sched_features.h"
649 static int sched_feat_open(struct inode *inode, struct file *filp)
651 filp->private_data = inode->i_private;
656 sched_feat_read(struct file *filp, char __user *ubuf,
657 size_t cnt, loff_t *ppos)
664 for (i = 0; sched_feat_names[i]; i++) {
665 len += strlen(sched_feat_names[i]);
669 buf = kmalloc(len + 2, GFP_KERNEL);
673 for (i = 0; sched_feat_names[i]; i++) {
674 if (sysctl_sched_features & (1UL << i))
675 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
677 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
680 r += sprintf(buf + r, "\n");
681 WARN_ON(r >= len + 2);
683 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
691 sched_feat_write(struct file *filp, const char __user *ubuf,
692 size_t cnt, loff_t *ppos)
702 if (copy_from_user(&buf, ubuf, cnt))
707 if (strncmp(buf, "NO_", 3) == 0) {
712 for (i = 0; sched_feat_names[i]; i++) {
713 int len = strlen(sched_feat_names[i]);
715 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
717 sysctl_sched_features &= ~(1UL << i);
719 sysctl_sched_features |= (1UL << i);
724 if (!sched_feat_names[i])
732 static struct file_operations sched_feat_fops = {
733 .open = sched_feat_open,
734 .read = sched_feat_read,
735 .write = sched_feat_write,
738 static __init int sched_init_debug(void)
740 debugfs_create_file("sched_features", 0644, NULL, NULL,
745 late_initcall(sched_init_debug);
749 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
752 * Number of tasks to iterate in a single balance run.
753 * Limited because this is done with IRQs disabled.
755 const_debug unsigned int sysctl_sched_nr_migrate = 32;
758 * period over which we measure -rt task cpu usage in us.
761 unsigned int sysctl_sched_rt_period = 1000000;
763 static __read_mostly int scheduler_running;
766 * part of the period that we allow rt tasks to run in us.
769 int sysctl_sched_rt_runtime = 950000;
771 static inline u64 global_rt_period(void)
773 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
776 static inline u64 global_rt_runtime(void)
778 if (sysctl_sched_rt_period < 0)
781 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
784 unsigned long long time_sync_thresh = 100000;
786 static DEFINE_PER_CPU(unsigned long long, time_offset);
787 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
790 * Global lock which we take every now and then to synchronize
791 * the CPUs time. This method is not warp-safe, but it's good
792 * enough to synchronize slowly diverging time sources and thus
793 * it's good enough for tracing:
795 static DEFINE_SPINLOCK(time_sync_lock);
796 static unsigned long long prev_global_time;
798 static unsigned long long __sync_cpu_clock(unsigned long long time, int cpu)
801 * We want this inlined, to not get tracer function calls
802 * in this critical section:
804 spin_acquire(&time_sync_lock.dep_map, 0, 0, _THIS_IP_);
805 __raw_spin_lock(&time_sync_lock.raw_lock);
807 if (time < prev_global_time) {
808 per_cpu(time_offset, cpu) += prev_global_time - time;
809 time = prev_global_time;
811 prev_global_time = time;
814 __raw_spin_unlock(&time_sync_lock.raw_lock);
815 spin_release(&time_sync_lock.dep_map, 1, _THIS_IP_);
820 static unsigned long long __cpu_clock(int cpu)
822 unsigned long long now;
825 * Only call sched_clock() if the scheduler has already been
826 * initialized (some code might call cpu_clock() very early):
828 if (unlikely(!scheduler_running))
831 now = sched_clock_cpu(cpu);
837 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
838 * clock constructed from sched_clock():
840 unsigned long long cpu_clock(int cpu)
842 unsigned long long prev_cpu_time, time, delta_time;
845 local_irq_save(flags);
846 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
847 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
848 delta_time = time-prev_cpu_time;
850 if (unlikely(delta_time > time_sync_thresh)) {
851 time = __sync_cpu_clock(time, cpu);
852 per_cpu(prev_cpu_time, cpu) = time;
854 local_irq_restore(flags);
858 EXPORT_SYMBOL_GPL(cpu_clock);
860 #ifndef prepare_arch_switch
861 # define prepare_arch_switch(next) do { } while (0)
863 #ifndef finish_arch_switch
864 # define finish_arch_switch(prev) do { } while (0)
867 static inline int task_current(struct rq *rq, struct task_struct *p)
869 return rq->curr == p;
872 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
873 static inline int task_running(struct rq *rq, struct task_struct *p)
875 return task_current(rq, p);
878 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
882 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
884 #ifdef CONFIG_DEBUG_SPINLOCK
885 /* this is a valid case when another task releases the spinlock */
886 rq->lock.owner = current;
889 * If we are tracking spinlock dependencies then we have to
890 * fix up the runqueue lock - which gets 'carried over' from
893 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
895 spin_unlock_irq(&rq->lock);
898 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
899 static inline int task_running(struct rq *rq, struct task_struct *p)
904 return task_current(rq, p);
908 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
912 * We can optimise this out completely for !SMP, because the
913 * SMP rebalancing from interrupt is the only thing that cares
918 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
919 spin_unlock_irq(&rq->lock);
921 spin_unlock(&rq->lock);
925 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
929 * After ->oncpu is cleared, the task can be moved to a different CPU.
930 * We must ensure this doesn't happen until the switch is completely
936 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
940 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
943 * __task_rq_lock - lock the runqueue a given task resides on.
944 * Must be called interrupts disabled.
946 static inline struct rq *__task_rq_lock(struct task_struct *p)
950 struct rq *rq = task_rq(p);
951 spin_lock(&rq->lock);
952 if (likely(rq == task_rq(p)))
954 spin_unlock(&rq->lock);
959 * task_rq_lock - lock the runqueue a given task resides on and disable
960 * interrupts. Note the ordering: we can safely lookup the task_rq without
961 * explicitly disabling preemption.
963 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
969 local_irq_save(*flags);
971 spin_lock(&rq->lock);
972 if (likely(rq == task_rq(p)))
974 spin_unlock_irqrestore(&rq->lock, *flags);
978 static void __task_rq_unlock(struct rq *rq)
981 spin_unlock(&rq->lock);
984 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
987 spin_unlock_irqrestore(&rq->lock, *flags);
991 * this_rq_lock - lock this runqueue and disable interrupts.
993 static struct rq *this_rq_lock(void)
1000 spin_lock(&rq->lock);
1005 static void __resched_task(struct task_struct *p, int tif_bit);
1007 static inline void resched_task(struct task_struct *p)
1009 __resched_task(p, TIF_NEED_RESCHED);
1012 #ifdef CONFIG_SCHED_HRTICK
1014 * Use HR-timers to deliver accurate preemption points.
1016 * Its all a bit involved since we cannot program an hrt while holding the
1017 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1020 * When we get rescheduled we reprogram the hrtick_timer outside of the
1023 static inline void resched_hrt(struct task_struct *p)
1025 __resched_task(p, TIF_HRTICK_RESCHED);
1028 static inline void resched_rq(struct rq *rq)
1030 unsigned long flags;
1032 spin_lock_irqsave(&rq->lock, flags);
1033 resched_task(rq->curr);
1034 spin_unlock_irqrestore(&rq->lock, flags);
1038 HRTICK_SET, /* re-programm hrtick_timer */
1039 HRTICK_RESET, /* not a new slice */
1040 HRTICK_BLOCK, /* stop hrtick operations */
1045 * - enabled by features
1046 * - hrtimer is actually high res
1048 static inline int hrtick_enabled(struct rq *rq)
1050 if (!sched_feat(HRTICK))
1052 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
1054 return hrtimer_is_hres_active(&rq->hrtick_timer);
1058 * Called to set the hrtick timer state.
1060 * called with rq->lock held and irqs disabled
1062 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1064 assert_spin_locked(&rq->lock);
1067 * preempt at: now + delay
1070 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1072 * indicate we need to program the timer
1074 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1076 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1079 * New slices are called from the schedule path and don't need a
1080 * forced reschedule.
1083 resched_hrt(rq->curr);
1086 static void hrtick_clear(struct rq *rq)
1088 if (hrtimer_active(&rq->hrtick_timer))
1089 hrtimer_cancel(&rq->hrtick_timer);
1093 * Update the timer from the possible pending state.
1095 static void hrtick_set(struct rq *rq)
1099 unsigned long flags;
1101 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1103 spin_lock_irqsave(&rq->lock, flags);
1104 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1105 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1106 time = rq->hrtick_expire;
1107 clear_thread_flag(TIF_HRTICK_RESCHED);
1108 spin_unlock_irqrestore(&rq->lock, flags);
1111 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1112 if (reset && !hrtimer_active(&rq->hrtick_timer))
1119 * High-resolution timer tick.
1120 * Runs from hardirq context with interrupts disabled.
1122 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1124 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1126 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1128 spin_lock(&rq->lock);
1129 update_rq_clock(rq);
1130 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1131 spin_unlock(&rq->lock);
1133 return HRTIMER_NORESTART;
1137 static void hotplug_hrtick_disable(int cpu)
1139 struct rq *rq = cpu_rq(cpu);
1140 unsigned long flags;
1142 spin_lock_irqsave(&rq->lock, flags);
1143 rq->hrtick_flags = 0;
1144 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1145 spin_unlock_irqrestore(&rq->lock, flags);
1150 static void hotplug_hrtick_enable(int cpu)
1152 struct rq *rq = cpu_rq(cpu);
1153 unsigned long flags;
1155 spin_lock_irqsave(&rq->lock, flags);
1156 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1157 spin_unlock_irqrestore(&rq->lock, flags);
1161 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1163 int cpu = (int)(long)hcpu;
1166 case CPU_UP_CANCELED:
1167 case CPU_UP_CANCELED_FROZEN:
1168 case CPU_DOWN_PREPARE:
1169 case CPU_DOWN_PREPARE_FROZEN:
1171 case CPU_DEAD_FROZEN:
1172 hotplug_hrtick_disable(cpu);
1175 case CPU_UP_PREPARE:
1176 case CPU_UP_PREPARE_FROZEN:
1177 case CPU_DOWN_FAILED:
1178 case CPU_DOWN_FAILED_FROZEN:
1180 case CPU_ONLINE_FROZEN:
1181 hotplug_hrtick_enable(cpu);
1188 static void init_hrtick(void)
1190 hotcpu_notifier(hotplug_hrtick, 0);
1192 #endif /* CONFIG_SMP */
1194 static void init_rq_hrtick(struct rq *rq)
1196 rq->hrtick_flags = 0;
1197 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1198 rq->hrtick_timer.function = hrtick;
1199 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1202 void hrtick_resched(void)
1205 unsigned long flags;
1207 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1210 local_irq_save(flags);
1211 rq = cpu_rq(smp_processor_id());
1213 local_irq_restore(flags);
1216 static inline void hrtick_clear(struct rq *rq)
1220 static inline void hrtick_set(struct rq *rq)
1224 static inline void init_rq_hrtick(struct rq *rq)
1228 void hrtick_resched(void)
1232 static inline void init_hrtick(void)
1238 * resched_task - mark a task 'to be rescheduled now'.
1240 * On UP this means the setting of the need_resched flag, on SMP it
1241 * might also involve a cross-CPU call to trigger the scheduler on
1246 #ifndef tsk_is_polling
1247 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1250 static void __resched_task(struct task_struct *p, int tif_bit)
1254 assert_spin_locked(&task_rq(p)->lock);
1256 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1259 set_tsk_thread_flag(p, tif_bit);
1262 if (cpu == smp_processor_id())
1265 /* NEED_RESCHED must be visible before we test polling */
1267 if (!tsk_is_polling(p))
1268 smp_send_reschedule(cpu);
1271 static void resched_cpu(int cpu)
1273 struct rq *rq = cpu_rq(cpu);
1274 unsigned long flags;
1276 if (!spin_trylock_irqsave(&rq->lock, flags))
1278 resched_task(cpu_curr(cpu));
1279 spin_unlock_irqrestore(&rq->lock, flags);
1284 * When add_timer_on() enqueues a timer into the timer wheel of an
1285 * idle CPU then this timer might expire before the next timer event
1286 * which is scheduled to wake up that CPU. In case of a completely
1287 * idle system the next event might even be infinite time into the
1288 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1289 * leaves the inner idle loop so the newly added timer is taken into
1290 * account when the CPU goes back to idle and evaluates the timer
1291 * wheel for the next timer event.
1293 void wake_up_idle_cpu(int cpu)
1295 struct rq *rq = cpu_rq(cpu);
1297 if (cpu == smp_processor_id())
1301 * This is safe, as this function is called with the timer
1302 * wheel base lock of (cpu) held. When the CPU is on the way
1303 * to idle and has not yet set rq->curr to idle then it will
1304 * be serialized on the timer wheel base lock and take the new
1305 * timer into account automatically.
1307 if (rq->curr != rq->idle)
1311 * We can set TIF_RESCHED on the idle task of the other CPU
1312 * lockless. The worst case is that the other CPU runs the
1313 * idle task through an additional NOOP schedule()
1315 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1317 /* NEED_RESCHED must be visible before we test polling */
1319 if (!tsk_is_polling(rq->idle))
1320 smp_send_reschedule(cpu);
1322 #endif /* CONFIG_NO_HZ */
1324 #else /* !CONFIG_SMP */
1325 static void __resched_task(struct task_struct *p, int tif_bit)
1327 assert_spin_locked(&task_rq(p)->lock);
1328 set_tsk_thread_flag(p, tif_bit);
1330 #endif /* CONFIG_SMP */
1332 #if BITS_PER_LONG == 32
1333 # define WMULT_CONST (~0UL)
1335 # define WMULT_CONST (1UL << 32)
1338 #define WMULT_SHIFT 32
1341 * Shift right and round:
1343 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1345 static unsigned long
1346 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1347 struct load_weight *lw)
1351 if (!lw->inv_weight) {
1352 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1355 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1359 tmp = (u64)delta_exec * weight;
1361 * Check whether we'd overflow the 64-bit multiplication:
1363 if (unlikely(tmp > WMULT_CONST))
1364 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1367 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1369 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1372 static inline unsigned long
1373 calc_delta_fair(unsigned long delta_exec, struct load_weight *lw)
1375 return calc_delta_mine(delta_exec, NICE_0_LOAD, lw);
1378 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1384 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1391 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1392 * of tasks with abnormal "nice" values across CPUs the contribution that
1393 * each task makes to its run queue's load is weighted according to its
1394 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1395 * scaled version of the new time slice allocation that they receive on time
1399 #define WEIGHT_IDLEPRIO 2
1400 #define WMULT_IDLEPRIO (1 << 31)
1403 * Nice levels are multiplicative, with a gentle 10% change for every
1404 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1405 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1406 * that remained on nice 0.
1408 * The "10% effect" is relative and cumulative: from _any_ nice level,
1409 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1410 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1411 * If a task goes up by ~10% and another task goes down by ~10% then
1412 * the relative distance between them is ~25%.)
1414 static const int prio_to_weight[40] = {
1415 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1416 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1417 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1418 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1419 /* 0 */ 1024, 820, 655, 526, 423,
1420 /* 5 */ 335, 272, 215, 172, 137,
1421 /* 10 */ 110, 87, 70, 56, 45,
1422 /* 15 */ 36, 29, 23, 18, 15,
1426 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1428 * In cases where the weight does not change often, we can use the
1429 * precalculated inverse to speed up arithmetics by turning divisions
1430 * into multiplications:
1432 static const u32 prio_to_wmult[40] = {
1433 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1434 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1435 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1436 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1437 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1438 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1439 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1440 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1443 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1446 * runqueue iterator, to support SMP load-balancing between different
1447 * scheduling classes, without having to expose their internal data
1448 * structures to the load-balancing proper:
1450 struct rq_iterator {
1452 struct task_struct *(*start)(void *);
1453 struct task_struct *(*next)(void *);
1457 static unsigned long
1458 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1459 unsigned long max_load_move, struct sched_domain *sd,
1460 enum cpu_idle_type idle, int *all_pinned,
1461 int *this_best_prio, struct rq_iterator *iterator);
1464 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1465 struct sched_domain *sd, enum cpu_idle_type idle,
1466 struct rq_iterator *iterator);
1469 #ifdef CONFIG_CGROUP_CPUACCT
1470 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1472 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1475 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1477 update_load_add(&rq->load, load);
1480 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1482 update_load_sub(&rq->load, load);
1486 static unsigned long source_load(int cpu, int type);
1487 static unsigned long target_load(int cpu, int type);
1488 static unsigned long cpu_avg_load_per_task(int cpu);
1489 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1492 #include "sched_stats.h"
1493 #include "sched_idletask.c"
1494 #include "sched_fair.c"
1495 #include "sched_rt.c"
1496 #ifdef CONFIG_SCHED_DEBUG
1497 # include "sched_debug.c"
1500 #define sched_class_highest (&rt_sched_class)
1501 #define for_each_class(class) \
1502 for (class = sched_class_highest; class; class = class->next)
1504 static inline void inc_load(struct rq *rq, const struct task_struct *p)
1506 update_load_add(&rq->load, p->se.load.weight);
1509 static inline void dec_load(struct rq *rq, const struct task_struct *p)
1511 update_load_sub(&rq->load, p->se.load.weight);
1514 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1520 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1526 static void set_load_weight(struct task_struct *p)
1528 if (task_has_rt_policy(p)) {
1529 p->se.load.weight = prio_to_weight[0] * 2;
1530 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1535 * SCHED_IDLE tasks get minimal weight:
1537 if (p->policy == SCHED_IDLE) {
1538 p->se.load.weight = WEIGHT_IDLEPRIO;
1539 p->se.load.inv_weight = WMULT_IDLEPRIO;
1543 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1544 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1547 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1549 sched_info_queued(p);
1550 p->sched_class->enqueue_task(rq, p, wakeup);
1554 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1556 p->sched_class->dequeue_task(rq, p, sleep);
1561 * __normal_prio - return the priority that is based on the static prio
1563 static inline int __normal_prio(struct task_struct *p)
1565 return p->static_prio;
1569 * Calculate the expected normal priority: i.e. priority
1570 * without taking RT-inheritance into account. Might be
1571 * boosted by interactivity modifiers. Changes upon fork,
1572 * setprio syscalls, and whenever the interactivity
1573 * estimator recalculates.
1575 static inline int normal_prio(struct task_struct *p)
1579 if (task_has_rt_policy(p))
1580 prio = MAX_RT_PRIO-1 - p->rt_priority;
1582 prio = __normal_prio(p);
1587 * Calculate the current priority, i.e. the priority
1588 * taken into account by the scheduler. This value might
1589 * be boosted by RT tasks, or might be boosted by
1590 * interactivity modifiers. Will be RT if the task got
1591 * RT-boosted. If not then it returns p->normal_prio.
1593 static int effective_prio(struct task_struct *p)
1595 p->normal_prio = normal_prio(p);
1597 * If we are RT tasks or we were boosted to RT priority,
1598 * keep the priority unchanged. Otherwise, update priority
1599 * to the normal priority:
1601 if (!rt_prio(p->prio))
1602 return p->normal_prio;
1607 * activate_task - move a task to the runqueue.
1609 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1611 if (task_contributes_to_load(p))
1612 rq->nr_uninterruptible--;
1614 enqueue_task(rq, p, wakeup);
1615 inc_nr_running(p, rq);
1619 * deactivate_task - remove a task from the runqueue.
1621 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1623 if (task_contributes_to_load(p))
1624 rq->nr_uninterruptible++;
1626 dequeue_task(rq, p, sleep);
1627 dec_nr_running(p, rq);
1631 * task_curr - is this task currently executing on a CPU?
1632 * @p: the task in question.
1634 inline int task_curr(const struct task_struct *p)
1636 return cpu_curr(task_cpu(p)) == p;
1639 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1641 set_task_rq(p, cpu);
1644 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1645 * successfuly executed on another CPU. We must ensure that updates of
1646 * per-task data have been completed by this moment.
1649 task_thread_info(p)->cpu = cpu;
1653 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1654 const struct sched_class *prev_class,
1655 int oldprio, int running)
1657 if (prev_class != p->sched_class) {
1658 if (prev_class->switched_from)
1659 prev_class->switched_from(rq, p, running);
1660 p->sched_class->switched_to(rq, p, running);
1662 p->sched_class->prio_changed(rq, p, oldprio, running);
1667 /* Used instead of source_load when we know the type == 0 */
1668 static unsigned long weighted_cpuload(const int cpu)
1670 return cpu_rq(cpu)->load.weight;
1674 * Is this task likely cache-hot:
1677 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1682 * Buddy candidates are cache hot:
1684 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1687 if (p->sched_class != &fair_sched_class)
1690 if (sysctl_sched_migration_cost == -1)
1692 if (sysctl_sched_migration_cost == 0)
1695 delta = now - p->se.exec_start;
1697 return delta < (s64)sysctl_sched_migration_cost;
1701 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1703 int old_cpu = task_cpu(p);
1704 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1705 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1706 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1709 clock_offset = old_rq->clock - new_rq->clock;
1711 #ifdef CONFIG_SCHEDSTATS
1712 if (p->se.wait_start)
1713 p->se.wait_start -= clock_offset;
1714 if (p->se.sleep_start)
1715 p->se.sleep_start -= clock_offset;
1716 if (p->se.block_start)
1717 p->se.block_start -= clock_offset;
1718 if (old_cpu != new_cpu) {
1719 schedstat_inc(p, se.nr_migrations);
1720 if (task_hot(p, old_rq->clock, NULL))
1721 schedstat_inc(p, se.nr_forced2_migrations);
1724 p->se.vruntime -= old_cfsrq->min_vruntime -
1725 new_cfsrq->min_vruntime;
1727 __set_task_cpu(p, new_cpu);
1730 struct migration_req {
1731 struct list_head list;
1733 struct task_struct *task;
1736 struct completion done;
1740 * The task's runqueue lock must be held.
1741 * Returns true if you have to wait for migration thread.
1744 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1746 struct rq *rq = task_rq(p);
1749 * If the task is not on a runqueue (and not running), then
1750 * it is sufficient to simply update the task's cpu field.
1752 if (!p->se.on_rq && !task_running(rq, p)) {
1753 set_task_cpu(p, dest_cpu);
1757 init_completion(&req->done);
1759 req->dest_cpu = dest_cpu;
1760 list_add(&req->list, &rq->migration_queue);
1766 * wait_task_inactive - wait for a thread to unschedule.
1768 * The caller must ensure that the task *will* unschedule sometime soon,
1769 * else this function might spin for a *long* time. This function can't
1770 * be called with interrupts off, or it may introduce deadlock with
1771 * smp_call_function() if an IPI is sent by the same process we are
1772 * waiting to become inactive.
1774 void wait_task_inactive(struct task_struct *p)
1776 unsigned long flags;
1782 * We do the initial early heuristics without holding
1783 * any task-queue locks at all. We'll only try to get
1784 * the runqueue lock when things look like they will
1790 * If the task is actively running on another CPU
1791 * still, just relax and busy-wait without holding
1794 * NOTE! Since we don't hold any locks, it's not
1795 * even sure that "rq" stays as the right runqueue!
1796 * But we don't care, since "task_running()" will
1797 * return false if the runqueue has changed and p
1798 * is actually now running somewhere else!
1800 while (task_running(rq, p))
1804 * Ok, time to look more closely! We need the rq
1805 * lock now, to be *sure*. If we're wrong, we'll
1806 * just go back and repeat.
1808 rq = task_rq_lock(p, &flags);
1809 running = task_running(rq, p);
1810 on_rq = p->se.on_rq;
1811 task_rq_unlock(rq, &flags);
1814 * Was it really running after all now that we
1815 * checked with the proper locks actually held?
1817 * Oops. Go back and try again..
1819 if (unlikely(running)) {
1825 * It's not enough that it's not actively running,
1826 * it must be off the runqueue _entirely_, and not
1829 * So if it wa still runnable (but just not actively
1830 * running right now), it's preempted, and we should
1831 * yield - it could be a while.
1833 if (unlikely(on_rq)) {
1834 schedule_timeout_uninterruptible(1);
1839 * Ahh, all good. It wasn't running, and it wasn't
1840 * runnable, which means that it will never become
1841 * running in the future either. We're all done!
1848 * kick_process - kick a running thread to enter/exit the kernel
1849 * @p: the to-be-kicked thread
1851 * Cause a process which is running on another CPU to enter
1852 * kernel-mode, without any delay. (to get signals handled.)
1854 * NOTE: this function doesnt have to take the runqueue lock,
1855 * because all it wants to ensure is that the remote task enters
1856 * the kernel. If the IPI races and the task has been migrated
1857 * to another CPU then no harm is done and the purpose has been
1860 void kick_process(struct task_struct *p)
1866 if ((cpu != smp_processor_id()) && task_curr(p))
1867 smp_send_reschedule(cpu);
1872 * Return a low guess at the load of a migration-source cpu weighted
1873 * according to the scheduling class and "nice" value.
1875 * We want to under-estimate the load of migration sources, to
1876 * balance conservatively.
1878 static unsigned long source_load(int cpu, int type)
1880 struct rq *rq = cpu_rq(cpu);
1881 unsigned long total = weighted_cpuload(cpu);
1886 return min(rq->cpu_load[type-1], total);
1890 * Return a high guess at the load of a migration-target cpu weighted
1891 * according to the scheduling class and "nice" value.
1893 static unsigned long target_load(int cpu, int type)
1895 struct rq *rq = cpu_rq(cpu);
1896 unsigned long total = weighted_cpuload(cpu);
1901 return max(rq->cpu_load[type-1], total);
1905 * Return the average load per task on the cpu's run queue
1907 static unsigned long cpu_avg_load_per_task(int cpu)
1909 struct rq *rq = cpu_rq(cpu);
1910 unsigned long total = weighted_cpuload(cpu);
1911 unsigned long n = rq->nr_running;
1913 return n ? total / n : SCHED_LOAD_SCALE;
1917 * find_idlest_group finds and returns the least busy CPU group within the
1920 static struct sched_group *
1921 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1923 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1924 unsigned long min_load = ULONG_MAX, this_load = 0;
1925 int load_idx = sd->forkexec_idx;
1926 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1929 unsigned long load, avg_load;
1933 /* Skip over this group if it has no CPUs allowed */
1934 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1937 local_group = cpu_isset(this_cpu, group->cpumask);
1939 /* Tally up the load of all CPUs in the group */
1942 for_each_cpu_mask(i, group->cpumask) {
1943 /* Bias balancing toward cpus of our domain */
1945 load = source_load(i, load_idx);
1947 load = target_load(i, load_idx);
1952 /* Adjust by relative CPU power of the group */
1953 avg_load = sg_div_cpu_power(group,
1954 avg_load * SCHED_LOAD_SCALE);
1957 this_load = avg_load;
1959 } else if (avg_load < min_load) {
1960 min_load = avg_load;
1963 } while (group = group->next, group != sd->groups);
1965 if (!idlest || 100*this_load < imbalance*min_load)
1971 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1974 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
1977 unsigned long load, min_load = ULONG_MAX;
1981 /* Traverse only the allowed CPUs */
1982 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
1984 for_each_cpu_mask(i, *tmp) {
1985 load = weighted_cpuload(i);
1987 if (load < min_load || (load == min_load && i == this_cpu)) {
1997 * sched_balance_self: balance the current task (running on cpu) in domains
1998 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2001 * Balance, ie. select the least loaded group.
2003 * Returns the target CPU number, or the same CPU if no balancing is needed.
2005 * preempt must be disabled.
2007 static int sched_balance_self(int cpu, int flag)
2009 struct task_struct *t = current;
2010 struct sched_domain *tmp, *sd = NULL;
2012 for_each_domain(cpu, tmp) {
2014 * If power savings logic is enabled for a domain, stop there.
2016 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2018 if (tmp->flags & flag)
2023 cpumask_t span, tmpmask;
2024 struct sched_group *group;
2025 int new_cpu, weight;
2027 if (!(sd->flags & flag)) {
2033 group = find_idlest_group(sd, t, cpu);
2039 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2040 if (new_cpu == -1 || new_cpu == cpu) {
2041 /* Now try balancing at a lower domain level of cpu */
2046 /* Now try balancing at a lower domain level of new_cpu */
2049 weight = cpus_weight(span);
2050 for_each_domain(cpu, tmp) {
2051 if (weight <= cpus_weight(tmp->span))
2053 if (tmp->flags & flag)
2056 /* while loop will break here if sd == NULL */
2062 #endif /* CONFIG_SMP */
2065 * try_to_wake_up - wake up a thread
2066 * @p: the to-be-woken-up thread
2067 * @state: the mask of task states that can be woken
2068 * @sync: do a synchronous wakeup?
2070 * Put it on the run-queue if it's not already there. The "current"
2071 * thread is always on the run-queue (except when the actual
2072 * re-schedule is in progress), and as such you're allowed to do
2073 * the simpler "current->state = TASK_RUNNING" to mark yourself
2074 * runnable without the overhead of this.
2076 * returns failure only if the task is already active.
2078 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2080 int cpu, orig_cpu, this_cpu, success = 0;
2081 unsigned long flags;
2085 if (!sched_feat(SYNC_WAKEUPS))
2089 rq = task_rq_lock(p, &flags);
2090 old_state = p->state;
2091 if (!(old_state & state))
2099 this_cpu = smp_processor_id();
2102 if (unlikely(task_running(rq, p)))
2105 cpu = p->sched_class->select_task_rq(p, sync);
2106 if (cpu != orig_cpu) {
2107 set_task_cpu(p, cpu);
2108 task_rq_unlock(rq, &flags);
2109 /* might preempt at this point */
2110 rq = task_rq_lock(p, &flags);
2111 old_state = p->state;
2112 if (!(old_state & state))
2117 this_cpu = smp_processor_id();
2121 #ifdef CONFIG_SCHEDSTATS
2122 schedstat_inc(rq, ttwu_count);
2123 if (cpu == this_cpu)
2124 schedstat_inc(rq, ttwu_local);
2126 struct sched_domain *sd;
2127 for_each_domain(this_cpu, sd) {
2128 if (cpu_isset(cpu, sd->span)) {
2129 schedstat_inc(sd, ttwu_wake_remote);
2134 #endif /* CONFIG_SCHEDSTATS */
2137 #endif /* CONFIG_SMP */
2138 schedstat_inc(p, se.nr_wakeups);
2140 schedstat_inc(p, se.nr_wakeups_sync);
2141 if (orig_cpu != cpu)
2142 schedstat_inc(p, se.nr_wakeups_migrate);
2143 if (cpu == this_cpu)
2144 schedstat_inc(p, se.nr_wakeups_local);
2146 schedstat_inc(p, se.nr_wakeups_remote);
2147 update_rq_clock(rq);
2148 activate_task(rq, p, 1);
2152 check_preempt_curr(rq, p);
2154 p->state = TASK_RUNNING;
2156 if (p->sched_class->task_wake_up)
2157 p->sched_class->task_wake_up(rq, p);
2160 task_rq_unlock(rq, &flags);
2165 int wake_up_process(struct task_struct *p)
2167 return try_to_wake_up(p, TASK_ALL, 0);
2169 EXPORT_SYMBOL(wake_up_process);
2171 int wake_up_state(struct task_struct *p, unsigned int state)
2173 return try_to_wake_up(p, state, 0);
2177 * Perform scheduler related setup for a newly forked process p.
2178 * p is forked by current.
2180 * __sched_fork() is basic setup used by init_idle() too:
2182 static void __sched_fork(struct task_struct *p)
2184 p->se.exec_start = 0;
2185 p->se.sum_exec_runtime = 0;
2186 p->se.prev_sum_exec_runtime = 0;
2187 p->se.last_wakeup = 0;
2188 p->se.avg_overlap = 0;
2190 #ifdef CONFIG_SCHEDSTATS
2191 p->se.wait_start = 0;
2192 p->se.sum_sleep_runtime = 0;
2193 p->se.sleep_start = 0;
2194 p->se.block_start = 0;
2195 p->se.sleep_max = 0;
2196 p->se.block_max = 0;
2198 p->se.slice_max = 0;
2202 INIT_LIST_HEAD(&p->rt.run_list);
2204 INIT_LIST_HEAD(&p->se.group_node);
2206 #ifdef CONFIG_PREEMPT_NOTIFIERS
2207 INIT_HLIST_HEAD(&p->preempt_notifiers);
2211 * We mark the process as running here, but have not actually
2212 * inserted it onto the runqueue yet. This guarantees that
2213 * nobody will actually run it, and a signal or other external
2214 * event cannot wake it up and insert it on the runqueue either.
2216 p->state = TASK_RUNNING;
2220 * fork()/clone()-time setup:
2222 void sched_fork(struct task_struct *p, int clone_flags)
2224 int cpu = get_cpu();
2229 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2231 set_task_cpu(p, cpu);
2234 * Make sure we do not leak PI boosting priority to the child:
2236 p->prio = current->normal_prio;
2237 if (!rt_prio(p->prio))
2238 p->sched_class = &fair_sched_class;
2240 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2241 if (likely(sched_info_on()))
2242 memset(&p->sched_info, 0, sizeof(p->sched_info));
2244 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2247 #ifdef CONFIG_PREEMPT
2248 /* Want to start with kernel preemption disabled. */
2249 task_thread_info(p)->preempt_count = 1;
2255 * wake_up_new_task - wake up a newly created task for the first time.
2257 * This function will do some initial scheduler statistics housekeeping
2258 * that must be done for every newly created context, then puts the task
2259 * on the runqueue and wakes it.
2261 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2263 unsigned long flags;
2266 rq = task_rq_lock(p, &flags);
2267 BUG_ON(p->state != TASK_RUNNING);
2268 update_rq_clock(rq);
2270 p->prio = effective_prio(p);
2272 if (!p->sched_class->task_new || !current->se.on_rq) {
2273 activate_task(rq, p, 0);
2276 * Let the scheduling class do new task startup
2277 * management (if any):
2279 p->sched_class->task_new(rq, p);
2280 inc_nr_running(p, rq);
2282 check_preempt_curr(rq, p);
2284 if (p->sched_class->task_wake_up)
2285 p->sched_class->task_wake_up(rq, p);
2287 task_rq_unlock(rq, &flags);
2290 #ifdef CONFIG_PREEMPT_NOTIFIERS
2293 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2294 * @notifier: notifier struct to register
2296 void preempt_notifier_register(struct preempt_notifier *notifier)
2298 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2300 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2303 * preempt_notifier_unregister - no longer interested in preemption notifications
2304 * @notifier: notifier struct to unregister
2306 * This is safe to call from within a preemption notifier.
2308 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2310 hlist_del(¬ifier->link);
2312 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2314 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2316 struct preempt_notifier *notifier;
2317 struct hlist_node *node;
2319 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2320 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2324 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2325 struct task_struct *next)
2327 struct preempt_notifier *notifier;
2328 struct hlist_node *node;
2330 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2331 notifier->ops->sched_out(notifier, next);
2334 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2336 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2341 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2342 struct task_struct *next)
2346 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2349 * prepare_task_switch - prepare to switch tasks
2350 * @rq: the runqueue preparing to switch
2351 * @prev: the current task that is being switched out
2352 * @next: the task we are going to switch to.
2354 * This is called with the rq lock held and interrupts off. It must
2355 * be paired with a subsequent finish_task_switch after the context
2358 * prepare_task_switch sets up locking and calls architecture specific
2362 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2363 struct task_struct *next)
2365 fire_sched_out_preempt_notifiers(prev, next);
2366 prepare_lock_switch(rq, next);
2367 prepare_arch_switch(next);
2371 * finish_task_switch - clean up after a task-switch
2372 * @rq: runqueue associated with task-switch
2373 * @prev: the thread we just switched away from.
2375 * finish_task_switch must be called after the context switch, paired
2376 * with a prepare_task_switch call before the context switch.
2377 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2378 * and do any other architecture-specific cleanup actions.
2380 * Note that we may have delayed dropping an mm in context_switch(). If
2381 * so, we finish that here outside of the runqueue lock. (Doing it
2382 * with the lock held can cause deadlocks; see schedule() for
2385 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2386 __releases(rq->lock)
2388 struct mm_struct *mm = rq->prev_mm;
2394 * A task struct has one reference for the use as "current".
2395 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2396 * schedule one last time. The schedule call will never return, and
2397 * the scheduled task must drop that reference.
2398 * The test for TASK_DEAD must occur while the runqueue locks are
2399 * still held, otherwise prev could be scheduled on another cpu, die
2400 * there before we look at prev->state, and then the reference would
2402 * Manfred Spraul <manfred@colorfullife.com>
2404 prev_state = prev->state;
2405 finish_arch_switch(prev);
2406 finish_lock_switch(rq, prev);
2408 if (current->sched_class->post_schedule)
2409 current->sched_class->post_schedule(rq);
2412 fire_sched_in_preempt_notifiers(current);
2415 if (unlikely(prev_state == TASK_DEAD)) {
2417 * Remove function-return probe instances associated with this
2418 * task and put them back on the free list.
2420 kprobe_flush_task(prev);
2421 put_task_struct(prev);
2426 * schedule_tail - first thing a freshly forked thread must call.
2427 * @prev: the thread we just switched away from.
2429 asmlinkage void schedule_tail(struct task_struct *prev)
2430 __releases(rq->lock)
2432 struct rq *rq = this_rq();
2434 finish_task_switch(rq, prev);
2435 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2436 /* In this case, finish_task_switch does not reenable preemption */
2439 if (current->set_child_tid)
2440 put_user(task_pid_vnr(current), current->set_child_tid);
2444 * context_switch - switch to the new MM and the new
2445 * thread's register state.
2448 context_switch(struct rq *rq, struct task_struct *prev,
2449 struct task_struct *next)
2451 struct mm_struct *mm, *oldmm;
2453 prepare_task_switch(rq, prev, next);
2455 oldmm = prev->active_mm;
2457 * For paravirt, this is coupled with an exit in switch_to to
2458 * combine the page table reload and the switch backend into
2461 arch_enter_lazy_cpu_mode();
2463 if (unlikely(!mm)) {
2464 next->active_mm = oldmm;
2465 atomic_inc(&oldmm->mm_count);
2466 enter_lazy_tlb(oldmm, next);
2468 switch_mm(oldmm, mm, next);
2470 if (unlikely(!prev->mm)) {
2471 prev->active_mm = NULL;
2472 rq->prev_mm = oldmm;
2475 * Since the runqueue lock will be released by the next
2476 * task (which is an invalid locking op but in the case
2477 * of the scheduler it's an obvious special-case), so we
2478 * do an early lockdep release here:
2480 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2481 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2484 /* Here we just switch the register state and the stack. */
2485 switch_to(prev, next, prev);
2489 * this_rq must be evaluated again because prev may have moved
2490 * CPUs since it called schedule(), thus the 'rq' on its stack
2491 * frame will be invalid.
2493 finish_task_switch(this_rq(), prev);
2497 * nr_running, nr_uninterruptible and nr_context_switches:
2499 * externally visible scheduler statistics: current number of runnable
2500 * threads, current number of uninterruptible-sleeping threads, total
2501 * number of context switches performed since bootup.
2503 unsigned long nr_running(void)
2505 unsigned long i, sum = 0;
2507 for_each_online_cpu(i)
2508 sum += cpu_rq(i)->nr_running;
2513 unsigned long nr_uninterruptible(void)
2515 unsigned long i, sum = 0;
2517 for_each_possible_cpu(i)
2518 sum += cpu_rq(i)->nr_uninterruptible;
2521 * Since we read the counters lockless, it might be slightly
2522 * inaccurate. Do not allow it to go below zero though:
2524 if (unlikely((long)sum < 0))
2530 unsigned long long nr_context_switches(void)
2533 unsigned long long sum = 0;
2535 for_each_possible_cpu(i)
2536 sum += cpu_rq(i)->nr_switches;
2541 unsigned long nr_iowait(void)
2543 unsigned long i, sum = 0;
2545 for_each_possible_cpu(i)
2546 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2551 unsigned long nr_active(void)
2553 unsigned long i, running = 0, uninterruptible = 0;
2555 for_each_online_cpu(i) {
2556 running += cpu_rq(i)->nr_running;
2557 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2560 if (unlikely((long)uninterruptible < 0))
2561 uninterruptible = 0;
2563 return running + uninterruptible;
2567 * Update rq->cpu_load[] statistics. This function is usually called every
2568 * scheduler tick (TICK_NSEC).
2570 static void update_cpu_load(struct rq *this_rq)
2572 unsigned long this_load = this_rq->load.weight;
2575 this_rq->nr_load_updates++;
2577 /* Update our load: */
2578 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2579 unsigned long old_load, new_load;
2581 /* scale is effectively 1 << i now, and >> i divides by scale */
2583 old_load = this_rq->cpu_load[i];
2584 new_load = this_load;
2586 * Round up the averaging division if load is increasing. This
2587 * prevents us from getting stuck on 9 if the load is 10, for
2590 if (new_load > old_load)
2591 new_load += scale-1;
2592 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2599 * double_rq_lock - safely lock two runqueues
2601 * Note this does not disable interrupts like task_rq_lock,
2602 * you need to do so manually before calling.
2604 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2605 __acquires(rq1->lock)
2606 __acquires(rq2->lock)
2608 BUG_ON(!irqs_disabled());
2610 spin_lock(&rq1->lock);
2611 __acquire(rq2->lock); /* Fake it out ;) */
2614 spin_lock(&rq1->lock);
2615 spin_lock(&rq2->lock);
2617 spin_lock(&rq2->lock);
2618 spin_lock(&rq1->lock);
2621 update_rq_clock(rq1);
2622 update_rq_clock(rq2);
2626 * double_rq_unlock - safely unlock two runqueues
2628 * Note this does not restore interrupts like task_rq_unlock,
2629 * you need to do so manually after calling.
2631 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2632 __releases(rq1->lock)
2633 __releases(rq2->lock)
2635 spin_unlock(&rq1->lock);
2637 spin_unlock(&rq2->lock);
2639 __release(rq2->lock);
2643 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2645 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2646 __releases(this_rq->lock)
2647 __acquires(busiest->lock)
2648 __acquires(this_rq->lock)
2652 if (unlikely(!irqs_disabled())) {
2653 /* printk() doesn't work good under rq->lock */
2654 spin_unlock(&this_rq->lock);
2657 if (unlikely(!spin_trylock(&busiest->lock))) {
2658 if (busiest < this_rq) {
2659 spin_unlock(&this_rq->lock);
2660 spin_lock(&busiest->lock);
2661 spin_lock(&this_rq->lock);
2664 spin_lock(&busiest->lock);
2670 * If dest_cpu is allowed for this process, migrate the task to it.
2671 * This is accomplished by forcing the cpu_allowed mask to only
2672 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2673 * the cpu_allowed mask is restored.
2675 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2677 struct migration_req req;
2678 unsigned long flags;
2681 rq = task_rq_lock(p, &flags);
2682 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2683 || unlikely(cpu_is_offline(dest_cpu)))
2686 /* force the process onto the specified CPU */
2687 if (migrate_task(p, dest_cpu, &req)) {
2688 /* Need to wait for migration thread (might exit: take ref). */
2689 struct task_struct *mt = rq->migration_thread;
2691 get_task_struct(mt);
2692 task_rq_unlock(rq, &flags);
2693 wake_up_process(mt);
2694 put_task_struct(mt);
2695 wait_for_completion(&req.done);
2700 task_rq_unlock(rq, &flags);
2704 * sched_exec - execve() is a valuable balancing opportunity, because at
2705 * this point the task has the smallest effective memory and cache footprint.
2707 void sched_exec(void)
2709 int new_cpu, this_cpu = get_cpu();
2710 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2712 if (new_cpu != this_cpu)
2713 sched_migrate_task(current, new_cpu);
2717 * pull_task - move a task from a remote runqueue to the local runqueue.
2718 * Both runqueues must be locked.
2720 static void pull_task(struct rq *src_rq, struct task_struct *p,
2721 struct rq *this_rq, int this_cpu)
2723 deactivate_task(src_rq, p, 0);
2724 set_task_cpu(p, this_cpu);
2725 activate_task(this_rq, p, 0);
2727 * Note that idle threads have a prio of MAX_PRIO, for this test
2728 * to be always true for them.
2730 check_preempt_curr(this_rq, p);
2734 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2737 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2738 struct sched_domain *sd, enum cpu_idle_type idle,
2742 * We do not migrate tasks that are:
2743 * 1) running (obviously), or
2744 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2745 * 3) are cache-hot on their current CPU.
2747 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2748 schedstat_inc(p, se.nr_failed_migrations_affine);
2753 if (task_running(rq, p)) {
2754 schedstat_inc(p, se.nr_failed_migrations_running);
2759 * Aggressive migration if:
2760 * 1) task is cache cold, or
2761 * 2) too many balance attempts have failed.
2764 if (!task_hot(p, rq->clock, sd) ||
2765 sd->nr_balance_failed > sd->cache_nice_tries) {
2766 #ifdef CONFIG_SCHEDSTATS
2767 if (task_hot(p, rq->clock, sd)) {
2768 schedstat_inc(sd, lb_hot_gained[idle]);
2769 schedstat_inc(p, se.nr_forced_migrations);
2775 if (task_hot(p, rq->clock, sd)) {
2776 schedstat_inc(p, se.nr_failed_migrations_hot);
2782 static unsigned long
2783 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2784 unsigned long max_load_move, struct sched_domain *sd,
2785 enum cpu_idle_type idle, int *all_pinned,
2786 int *this_best_prio, struct rq_iterator *iterator)
2788 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2789 struct task_struct *p;
2790 long rem_load_move = max_load_move;
2792 if (max_load_move == 0)
2798 * Start the load-balancing iterator:
2800 p = iterator->start(iterator->arg);
2802 if (!p || loops++ > sysctl_sched_nr_migrate)
2805 * To help distribute high priority tasks across CPUs we don't
2806 * skip a task if it will be the highest priority task (i.e. smallest
2807 * prio value) on its new queue regardless of its load weight
2809 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2810 SCHED_LOAD_SCALE_FUZZ;
2811 if ((skip_for_load && p->prio >= *this_best_prio) ||
2812 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2813 p = iterator->next(iterator->arg);
2817 pull_task(busiest, p, this_rq, this_cpu);
2819 rem_load_move -= p->se.load.weight;
2822 * We only want to steal up to the prescribed amount of weighted load.
2824 if (rem_load_move > 0) {
2825 if (p->prio < *this_best_prio)
2826 *this_best_prio = p->prio;
2827 p = iterator->next(iterator->arg);
2832 * Right now, this is one of only two places pull_task() is called,
2833 * so we can safely collect pull_task() stats here rather than
2834 * inside pull_task().
2836 schedstat_add(sd, lb_gained[idle], pulled);
2839 *all_pinned = pinned;
2841 return max_load_move - rem_load_move;
2845 * move_tasks tries to move up to max_load_move weighted load from busiest to
2846 * this_rq, as part of a balancing operation within domain "sd".
2847 * Returns 1 if successful and 0 otherwise.
2849 * Called with both runqueues locked.
2851 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2852 unsigned long max_load_move,
2853 struct sched_domain *sd, enum cpu_idle_type idle,
2856 const struct sched_class *class = sched_class_highest;
2857 unsigned long total_load_moved = 0;
2858 int this_best_prio = this_rq->curr->prio;
2862 class->load_balance(this_rq, this_cpu, busiest,
2863 max_load_move - total_load_moved,
2864 sd, idle, all_pinned, &this_best_prio);
2865 class = class->next;
2866 } while (class && max_load_move > total_load_moved);
2868 return total_load_moved > 0;
2872 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2873 struct sched_domain *sd, enum cpu_idle_type idle,
2874 struct rq_iterator *iterator)
2876 struct task_struct *p = iterator->start(iterator->arg);
2880 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2881 pull_task(busiest, p, this_rq, this_cpu);
2883 * Right now, this is only the second place pull_task()
2884 * is called, so we can safely collect pull_task()
2885 * stats here rather than inside pull_task().
2887 schedstat_inc(sd, lb_gained[idle]);
2891 p = iterator->next(iterator->arg);
2898 * move_one_task tries to move exactly one task from busiest to this_rq, as
2899 * part of active balancing operations within "domain".
2900 * Returns 1 if successful and 0 otherwise.
2902 * Called with both runqueues locked.
2904 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2905 struct sched_domain *sd, enum cpu_idle_type idle)
2907 const struct sched_class *class;
2909 for (class = sched_class_highest; class; class = class->next)
2910 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2917 * find_busiest_group finds and returns the busiest CPU group within the
2918 * domain. It calculates and returns the amount of weighted load which
2919 * should be moved to restore balance via the imbalance parameter.
2921 static struct sched_group *
2922 find_busiest_group(struct sched_domain *sd, int this_cpu,
2923 unsigned long *imbalance, enum cpu_idle_type idle,
2924 int *sd_idle, const cpumask_t *cpus, int *balance)
2926 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2927 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2928 unsigned long max_pull;
2929 unsigned long busiest_load_per_task, busiest_nr_running;
2930 unsigned long this_load_per_task, this_nr_running;
2931 int load_idx, group_imb = 0;
2932 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2933 int power_savings_balance = 1;
2934 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2935 unsigned long min_nr_running = ULONG_MAX;
2936 struct sched_group *group_min = NULL, *group_leader = NULL;
2939 max_load = this_load = total_load = total_pwr = 0;
2940 busiest_load_per_task = busiest_nr_running = 0;
2941 this_load_per_task = this_nr_running = 0;
2942 if (idle == CPU_NOT_IDLE)
2943 load_idx = sd->busy_idx;
2944 else if (idle == CPU_NEWLY_IDLE)
2945 load_idx = sd->newidle_idx;
2947 load_idx = sd->idle_idx;
2950 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2953 int __group_imb = 0;
2954 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2955 unsigned long sum_nr_running, sum_weighted_load;
2957 local_group = cpu_isset(this_cpu, group->cpumask);
2960 balance_cpu = first_cpu(group->cpumask);
2962 /* Tally up the load of all CPUs in the group */
2963 sum_weighted_load = sum_nr_running = avg_load = 0;
2965 min_cpu_load = ~0UL;
2967 for_each_cpu_mask(i, group->cpumask) {
2970 if (!cpu_isset(i, *cpus))
2975 if (*sd_idle && rq->nr_running)
2978 /* Bias balancing toward cpus of our domain */
2980 if (idle_cpu(i) && !first_idle_cpu) {
2985 load = target_load(i, load_idx);
2987 load = source_load(i, load_idx);
2988 if (load > max_cpu_load)
2989 max_cpu_load = load;
2990 if (min_cpu_load > load)
2991 min_cpu_load = load;
2995 sum_nr_running += rq->nr_running;
2996 sum_weighted_load += weighted_cpuload(i);
3000 * First idle cpu or the first cpu(busiest) in this sched group
3001 * is eligible for doing load balancing at this and above
3002 * domains. In the newly idle case, we will allow all the cpu's
3003 * to do the newly idle load balance.
3005 if (idle != CPU_NEWLY_IDLE && local_group &&
3006 balance_cpu != this_cpu && balance) {
3011 total_load += avg_load;
3012 total_pwr += group->__cpu_power;
3014 /* Adjust by relative CPU power of the group */
3015 avg_load = sg_div_cpu_power(group,
3016 avg_load * SCHED_LOAD_SCALE);
3018 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3021 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3024 this_load = avg_load;
3026 this_nr_running = sum_nr_running;
3027 this_load_per_task = sum_weighted_load;
3028 } else if (avg_load > max_load &&
3029 (sum_nr_running > group_capacity || __group_imb)) {
3030 max_load = avg_load;
3032 busiest_nr_running = sum_nr_running;
3033 busiest_load_per_task = sum_weighted_load;
3034 group_imb = __group_imb;
3037 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3039 * Busy processors will not participate in power savings
3042 if (idle == CPU_NOT_IDLE ||
3043 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3047 * If the local group is idle or completely loaded
3048 * no need to do power savings balance at this domain
3050 if (local_group && (this_nr_running >= group_capacity ||
3052 power_savings_balance = 0;
3055 * If a group is already running at full capacity or idle,
3056 * don't include that group in power savings calculations
3058 if (!power_savings_balance || sum_nr_running >= group_capacity
3063 * Calculate the group which has the least non-idle load.
3064 * This is the group from where we need to pick up the load
3067 if ((sum_nr_running < min_nr_running) ||
3068 (sum_nr_running == min_nr_running &&
3069 first_cpu(group->cpumask) <
3070 first_cpu(group_min->cpumask))) {
3072 min_nr_running = sum_nr_running;
3073 min_load_per_task = sum_weighted_load /
3078 * Calculate the group which is almost near its
3079 * capacity but still has some space to pick up some load
3080 * from other group and save more power
3082 if (sum_nr_running <= group_capacity - 1) {
3083 if (sum_nr_running > leader_nr_running ||
3084 (sum_nr_running == leader_nr_running &&
3085 first_cpu(group->cpumask) >
3086 first_cpu(group_leader->cpumask))) {
3087 group_leader = group;
3088 leader_nr_running = sum_nr_running;
3093 group = group->next;
3094 } while (group != sd->groups);
3096 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3099 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3101 if (this_load >= avg_load ||
3102 100*max_load <= sd->imbalance_pct*this_load)
3105 busiest_load_per_task /= busiest_nr_running;
3107 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3110 * We're trying to get all the cpus to the average_load, so we don't
3111 * want to push ourselves above the average load, nor do we wish to
3112 * reduce the max loaded cpu below the average load, as either of these
3113 * actions would just result in more rebalancing later, and ping-pong
3114 * tasks around. Thus we look for the minimum possible imbalance.
3115 * Negative imbalances (*we* are more loaded than anyone else) will
3116 * be counted as no imbalance for these purposes -- we can't fix that
3117 * by pulling tasks to us. Be careful of negative numbers as they'll
3118 * appear as very large values with unsigned longs.
3120 if (max_load <= busiest_load_per_task)
3124 * In the presence of smp nice balancing, certain scenarios can have
3125 * max load less than avg load(as we skip the groups at or below
3126 * its cpu_power, while calculating max_load..)
3128 if (max_load < avg_load) {
3130 goto small_imbalance;
3133 /* Don't want to pull so many tasks that a group would go idle */
3134 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3136 /* How much load to actually move to equalise the imbalance */
3137 *imbalance = min(max_pull * busiest->__cpu_power,
3138 (avg_load - this_load) * this->__cpu_power)
3142 * if *imbalance is less than the average load per runnable task
3143 * there is no gaurantee that any tasks will be moved so we'll have
3144 * a think about bumping its value to force at least one task to be
3147 if (*imbalance < busiest_load_per_task) {
3148 unsigned long tmp, pwr_now, pwr_move;
3152 pwr_move = pwr_now = 0;
3154 if (this_nr_running) {
3155 this_load_per_task /= this_nr_running;
3156 if (busiest_load_per_task > this_load_per_task)
3159 this_load_per_task = SCHED_LOAD_SCALE;
3161 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3162 busiest_load_per_task * imbn) {
3163 *imbalance = busiest_load_per_task;
3168 * OK, we don't have enough imbalance to justify moving tasks,
3169 * however we may be able to increase total CPU power used by
3173 pwr_now += busiest->__cpu_power *
3174 min(busiest_load_per_task, max_load);
3175 pwr_now += this->__cpu_power *
3176 min(this_load_per_task, this_load);
3177 pwr_now /= SCHED_LOAD_SCALE;
3179 /* Amount of load we'd subtract */
3180 tmp = sg_div_cpu_power(busiest,
3181 busiest_load_per_task * SCHED_LOAD_SCALE);
3183 pwr_move += busiest->__cpu_power *
3184 min(busiest_load_per_task, max_load - tmp);
3186 /* Amount of load we'd add */
3187 if (max_load * busiest->__cpu_power <
3188 busiest_load_per_task * SCHED_LOAD_SCALE)
3189 tmp = sg_div_cpu_power(this,
3190 max_load * busiest->__cpu_power);
3192 tmp = sg_div_cpu_power(this,
3193 busiest_load_per_task * SCHED_LOAD_SCALE);
3194 pwr_move += this->__cpu_power *
3195 min(this_load_per_task, this_load + tmp);
3196 pwr_move /= SCHED_LOAD_SCALE;
3198 /* Move if we gain throughput */
3199 if (pwr_move > pwr_now)
3200 *imbalance = busiest_load_per_task;
3206 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3207 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3210 if (this == group_leader && group_leader != group_min) {
3211 *imbalance = min_load_per_task;
3221 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3224 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3225 unsigned long imbalance, const cpumask_t *cpus)
3227 struct rq *busiest = NULL, *rq;
3228 unsigned long max_load = 0;
3231 for_each_cpu_mask(i, group->cpumask) {
3234 if (!cpu_isset(i, *cpus))
3238 wl = weighted_cpuload(i);
3240 if (rq->nr_running == 1 && wl > imbalance)
3243 if (wl > max_load) {
3253 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3254 * so long as it is large enough.
3256 #define MAX_PINNED_INTERVAL 512
3259 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3260 * tasks if there is an imbalance.
3262 static int load_balance(int this_cpu, struct rq *this_rq,
3263 struct sched_domain *sd, enum cpu_idle_type idle,
3264 int *balance, cpumask_t *cpus)
3266 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3267 struct sched_group *group;
3268 unsigned long imbalance;
3270 unsigned long flags;
3275 * When power savings policy is enabled for the parent domain, idle
3276 * sibling can pick up load irrespective of busy siblings. In this case,
3277 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3278 * portraying it as CPU_NOT_IDLE.
3280 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3281 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3284 schedstat_inc(sd, lb_count[idle]);
3287 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3294 schedstat_inc(sd, lb_nobusyg[idle]);
3298 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3300 schedstat_inc(sd, lb_nobusyq[idle]);
3304 BUG_ON(busiest == this_rq);
3306 schedstat_add(sd, lb_imbalance[idle], imbalance);
3309 if (busiest->nr_running > 1) {
3311 * Attempt to move tasks. If find_busiest_group has found
3312 * an imbalance but busiest->nr_running <= 1, the group is
3313 * still unbalanced. ld_moved simply stays zero, so it is
3314 * correctly treated as an imbalance.
3316 local_irq_save(flags);
3317 double_rq_lock(this_rq, busiest);
3318 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3319 imbalance, sd, idle, &all_pinned);
3320 double_rq_unlock(this_rq, busiest);
3321 local_irq_restore(flags);
3324 * some other cpu did the load balance for us.
3326 if (ld_moved && this_cpu != smp_processor_id())
3327 resched_cpu(this_cpu);
3329 /* All tasks on this runqueue were pinned by CPU affinity */
3330 if (unlikely(all_pinned)) {
3331 cpu_clear(cpu_of(busiest), *cpus);
3332 if (!cpus_empty(*cpus))
3339 schedstat_inc(sd, lb_failed[idle]);
3340 sd->nr_balance_failed++;
3342 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3344 spin_lock_irqsave(&busiest->lock, flags);
3346 /* don't kick the migration_thread, if the curr
3347 * task on busiest cpu can't be moved to this_cpu
3349 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3350 spin_unlock_irqrestore(&busiest->lock, flags);
3352 goto out_one_pinned;
3355 if (!busiest->active_balance) {
3356 busiest->active_balance = 1;
3357 busiest->push_cpu = this_cpu;
3360 spin_unlock_irqrestore(&busiest->lock, flags);
3362 wake_up_process(busiest->migration_thread);
3365 * We've kicked active balancing, reset the failure
3368 sd->nr_balance_failed = sd->cache_nice_tries+1;
3371 sd->nr_balance_failed = 0;
3373 if (likely(!active_balance)) {
3374 /* We were unbalanced, so reset the balancing interval */
3375 sd->balance_interval = sd->min_interval;
3378 * If we've begun active balancing, start to back off. This
3379 * case may not be covered by the all_pinned logic if there
3380 * is only 1 task on the busy runqueue (because we don't call
3383 if (sd->balance_interval < sd->max_interval)
3384 sd->balance_interval *= 2;
3387 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3388 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3393 schedstat_inc(sd, lb_balanced[idle]);
3395 sd->nr_balance_failed = 0;
3398 /* tune up the balancing interval */
3399 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3400 (sd->balance_interval < sd->max_interval))
3401 sd->balance_interval *= 2;
3403 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3404 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3410 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3411 * tasks if there is an imbalance.
3413 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3414 * this_rq is locked.
3417 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3420 struct sched_group *group;
3421 struct rq *busiest = NULL;
3422 unsigned long imbalance;
3430 * When power savings policy is enabled for the parent domain, idle
3431 * sibling can pick up load irrespective of busy siblings. In this case,
3432 * let the state of idle sibling percolate up as IDLE, instead of
3433 * portraying it as CPU_NOT_IDLE.
3435 if (sd->flags & SD_SHARE_CPUPOWER &&
3436 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3439 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3441 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3442 &sd_idle, cpus, NULL);
3444 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3448 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3450 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3454 BUG_ON(busiest == this_rq);
3456 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3459 if (busiest->nr_running > 1) {
3460 /* Attempt to move tasks */
3461 double_lock_balance(this_rq, busiest);
3462 /* this_rq->clock is already updated */
3463 update_rq_clock(busiest);
3464 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3465 imbalance, sd, CPU_NEWLY_IDLE,
3467 spin_unlock(&busiest->lock);
3469 if (unlikely(all_pinned)) {
3470 cpu_clear(cpu_of(busiest), *cpus);
3471 if (!cpus_empty(*cpus))
3477 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3478 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3479 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3482 sd->nr_balance_failed = 0;
3487 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3488 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3489 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3491 sd->nr_balance_failed = 0;
3497 * idle_balance is called by schedule() if this_cpu is about to become
3498 * idle. Attempts to pull tasks from other CPUs.
3500 static void idle_balance(int this_cpu, struct rq *this_rq)
3502 struct sched_domain *sd;
3503 int pulled_task = -1;
3504 unsigned long next_balance = jiffies + HZ;
3507 for_each_domain(this_cpu, sd) {
3508 unsigned long interval;
3510 if (!(sd->flags & SD_LOAD_BALANCE))
3513 if (sd->flags & SD_BALANCE_NEWIDLE)
3514 /* If we've pulled tasks over stop searching: */
3515 pulled_task = load_balance_newidle(this_cpu, this_rq,
3518 interval = msecs_to_jiffies(sd->balance_interval);
3519 if (time_after(next_balance, sd->last_balance + interval))
3520 next_balance = sd->last_balance + interval;
3524 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3526 * We are going idle. next_balance may be set based on
3527 * a busy processor. So reset next_balance.
3529 this_rq->next_balance = next_balance;
3534 * active_load_balance is run by migration threads. It pushes running tasks
3535 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3536 * running on each physical CPU where possible, and avoids physical /
3537 * logical imbalances.
3539 * Called with busiest_rq locked.
3541 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3543 int target_cpu = busiest_rq->push_cpu;
3544 struct sched_domain *sd;
3545 struct rq *target_rq;
3547 /* Is there any task to move? */
3548 if (busiest_rq->nr_running <= 1)
3551 target_rq = cpu_rq(target_cpu);
3554 * This condition is "impossible", if it occurs
3555 * we need to fix it. Originally reported by
3556 * Bjorn Helgaas on a 128-cpu setup.
3558 BUG_ON(busiest_rq == target_rq);
3560 /* move a task from busiest_rq to target_rq */
3561 double_lock_balance(busiest_rq, target_rq);
3562 update_rq_clock(busiest_rq);
3563 update_rq_clock(target_rq);
3565 /* Search for an sd spanning us and the target CPU. */
3566 for_each_domain(target_cpu, sd) {
3567 if ((sd->flags & SD_LOAD_BALANCE) &&
3568 cpu_isset(busiest_cpu, sd->span))
3573 schedstat_inc(sd, alb_count);
3575 if (move_one_task(target_rq, target_cpu, busiest_rq,
3577 schedstat_inc(sd, alb_pushed);
3579 schedstat_inc(sd, alb_failed);
3581 spin_unlock(&target_rq->lock);
3586 atomic_t load_balancer;
3588 } nohz ____cacheline_aligned = {
3589 .load_balancer = ATOMIC_INIT(-1),
3590 .cpu_mask = CPU_MASK_NONE,
3594 * This routine will try to nominate the ilb (idle load balancing)
3595 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3596 * load balancing on behalf of all those cpus. If all the cpus in the system
3597 * go into this tickless mode, then there will be no ilb owner (as there is
3598 * no need for one) and all the cpus will sleep till the next wakeup event
3601 * For the ilb owner, tick is not stopped. And this tick will be used
3602 * for idle load balancing. ilb owner will still be part of
3605 * While stopping the tick, this cpu will become the ilb owner if there
3606 * is no other owner. And will be the owner till that cpu becomes busy
3607 * or if all cpus in the system stop their ticks at which point
3608 * there is no need for ilb owner.
3610 * When the ilb owner becomes busy, it nominates another owner, during the
3611 * next busy scheduler_tick()
3613 int select_nohz_load_balancer(int stop_tick)
3615 int cpu = smp_processor_id();
3618 cpu_set(cpu, nohz.cpu_mask);
3619 cpu_rq(cpu)->in_nohz_recently = 1;
3622 * If we are going offline and still the leader, give up!
3624 if (cpu_is_offline(cpu) &&
3625 atomic_read(&nohz.load_balancer) == cpu) {
3626 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3631 /* time for ilb owner also to sleep */
3632 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3633 if (atomic_read(&nohz.load_balancer) == cpu)
3634 atomic_set(&nohz.load_balancer, -1);
3638 if (atomic_read(&nohz.load_balancer) == -1) {
3639 /* make me the ilb owner */
3640 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3642 } else if (atomic_read(&nohz.load_balancer) == cpu)
3645 if (!cpu_isset(cpu, nohz.cpu_mask))
3648 cpu_clear(cpu, nohz.cpu_mask);
3650 if (atomic_read(&nohz.load_balancer) == cpu)
3651 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3658 static DEFINE_SPINLOCK(balancing);
3661 * It checks each scheduling domain to see if it is due to be balanced,
3662 * and initiates a balancing operation if so.
3664 * Balancing parameters are set up in arch_init_sched_domains.
3666 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3669 struct rq *rq = cpu_rq(cpu);
3670 unsigned long interval;
3671 struct sched_domain *sd;
3672 /* Earliest time when we have to do rebalance again */
3673 unsigned long next_balance = jiffies + 60*HZ;
3674 int update_next_balance = 0;
3678 for_each_domain(cpu, sd) {
3679 if (!(sd->flags & SD_LOAD_BALANCE))
3682 interval = sd->balance_interval;
3683 if (idle != CPU_IDLE)
3684 interval *= sd->busy_factor;
3686 /* scale ms to jiffies */
3687 interval = msecs_to_jiffies(interval);
3688 if (unlikely(!interval))
3690 if (interval > HZ*NR_CPUS/10)
3691 interval = HZ*NR_CPUS/10;
3693 need_serialize = sd->flags & SD_SERIALIZE;
3695 if (need_serialize) {
3696 if (!spin_trylock(&balancing))
3700 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3701 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3703 * We've pulled tasks over so either we're no
3704 * longer idle, or one of our SMT siblings is
3707 idle = CPU_NOT_IDLE;
3709 sd->last_balance = jiffies;
3712 spin_unlock(&balancing);
3714 if (time_after(next_balance, sd->last_balance + interval)) {
3715 next_balance = sd->last_balance + interval;
3716 update_next_balance = 1;
3720 * Stop the load balance at this level. There is another
3721 * CPU in our sched group which is doing load balancing more
3729 * next_balance will be updated only when there is a need.
3730 * When the cpu is attached to null domain for ex, it will not be
3733 if (likely(update_next_balance))
3734 rq->next_balance = next_balance;
3738 * run_rebalance_domains is triggered when needed from the scheduler tick.
3739 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3740 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3742 static void run_rebalance_domains(struct softirq_action *h)
3744 int this_cpu = smp_processor_id();
3745 struct rq *this_rq = cpu_rq(this_cpu);
3746 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3747 CPU_IDLE : CPU_NOT_IDLE;
3749 rebalance_domains(this_cpu, idle);
3753 * If this cpu is the owner for idle load balancing, then do the
3754 * balancing on behalf of the other idle cpus whose ticks are
3757 if (this_rq->idle_at_tick &&
3758 atomic_read(&nohz.load_balancer) == this_cpu) {
3759 cpumask_t cpus = nohz.cpu_mask;
3763 cpu_clear(this_cpu, cpus);
3764 for_each_cpu_mask(balance_cpu, cpus) {
3766 * If this cpu gets work to do, stop the load balancing
3767 * work being done for other cpus. Next load
3768 * balancing owner will pick it up.
3773 rebalance_domains(balance_cpu, CPU_IDLE);
3775 rq = cpu_rq(balance_cpu);
3776 if (time_after(this_rq->next_balance, rq->next_balance))
3777 this_rq->next_balance = rq->next_balance;
3784 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3786 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3787 * idle load balancing owner or decide to stop the periodic load balancing,
3788 * if the whole system is idle.
3790 static inline void trigger_load_balance(struct rq *rq, int cpu)
3794 * If we were in the nohz mode recently and busy at the current
3795 * scheduler tick, then check if we need to nominate new idle
3798 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3799 rq->in_nohz_recently = 0;
3801 if (atomic_read(&nohz.load_balancer) == cpu) {
3802 cpu_clear(cpu, nohz.cpu_mask);
3803 atomic_set(&nohz.load_balancer, -1);
3806 if (atomic_read(&nohz.load_balancer) == -1) {
3808 * simple selection for now: Nominate the
3809 * first cpu in the nohz list to be the next
3812 * TBD: Traverse the sched domains and nominate
3813 * the nearest cpu in the nohz.cpu_mask.
3815 int ilb = first_cpu(nohz.cpu_mask);
3817 if (ilb < nr_cpu_ids)
3823 * If this cpu is idle and doing idle load balancing for all the
3824 * cpus with ticks stopped, is it time for that to stop?
3826 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3827 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3833 * If this cpu is idle and the idle load balancing is done by
3834 * someone else, then no need raise the SCHED_SOFTIRQ
3836 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3837 cpu_isset(cpu, nohz.cpu_mask))
3840 if (time_after_eq(jiffies, rq->next_balance))
3841 raise_softirq(SCHED_SOFTIRQ);
3844 #else /* CONFIG_SMP */
3847 * on UP we do not need to balance between CPUs:
3849 static inline void idle_balance(int cpu, struct rq *rq)
3855 DEFINE_PER_CPU(struct kernel_stat, kstat);
3857 EXPORT_PER_CPU_SYMBOL(kstat);
3860 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3861 * that have not yet been banked in case the task is currently running.
3863 unsigned long long task_sched_runtime(struct task_struct *p)
3865 unsigned long flags;
3869 rq = task_rq_lock(p, &flags);
3870 ns = p->se.sum_exec_runtime;
3871 if (task_current(rq, p)) {
3872 update_rq_clock(rq);
3873 delta_exec = rq->clock - p->se.exec_start;
3874 if ((s64)delta_exec > 0)
3877 task_rq_unlock(rq, &flags);
3883 * Account user cpu time to a process.
3884 * @p: the process that the cpu time gets accounted to
3885 * @cputime: the cpu time spent in user space since the last update
3887 void account_user_time(struct task_struct *p, cputime_t cputime)
3889 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3892 p->utime = cputime_add(p->utime, cputime);
3894 /* Add user time to cpustat. */
3895 tmp = cputime_to_cputime64(cputime);
3896 if (TASK_NICE(p) > 0)
3897 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3899 cpustat->user = cputime64_add(cpustat->user, tmp);
3903 * Account guest cpu time to a process.
3904 * @p: the process that the cpu time gets accounted to
3905 * @cputime: the cpu time spent in virtual machine since the last update
3907 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3910 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3912 tmp = cputime_to_cputime64(cputime);
3914 p->utime = cputime_add(p->utime, cputime);
3915 p->gtime = cputime_add(p->gtime, cputime);
3917 cpustat->user = cputime64_add(cpustat->user, tmp);
3918 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3922 * Account scaled user cpu time to a process.
3923 * @p: the process that the cpu time gets accounted to
3924 * @cputime: the cpu time spent in user space since the last update
3926 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3928 p->utimescaled = cputime_add(p->utimescaled, cputime);
3932 * Account system cpu time to a process.
3933 * @p: the process that the cpu time gets accounted to
3934 * @hardirq_offset: the offset to subtract from hardirq_count()
3935 * @cputime: the cpu time spent in kernel space since the last update
3937 void account_system_time(struct task_struct *p, int hardirq_offset,
3940 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3941 struct rq *rq = this_rq();
3944 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3945 account_guest_time(p, cputime);
3949 p->stime = cputime_add(p->stime, cputime);
3951 /* Add system time to cpustat. */
3952 tmp = cputime_to_cputime64(cputime);
3953 if (hardirq_count() - hardirq_offset)
3954 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3955 else if (softirq_count())
3956 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3957 else if (p != rq->idle)
3958 cpustat->system = cputime64_add(cpustat->system, tmp);
3959 else if (atomic_read(&rq->nr_iowait) > 0)
3960 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3962 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3963 /* Account for system time used */
3964 acct_update_integrals(p);
3968 * Account scaled system cpu time to a process.
3969 * @p: the process that the cpu time gets accounted to
3970 * @hardirq_offset: the offset to subtract from hardirq_count()
3971 * @cputime: the cpu time spent in kernel space since the last update
3973 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3975 p->stimescaled = cputime_add(p->stimescaled, cputime);
3979 * Account for involuntary wait time.
3980 * @p: the process from which the cpu time has been stolen
3981 * @steal: the cpu time spent in involuntary wait
3983 void account_steal_time(struct task_struct *p, cputime_t steal)
3985 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3986 cputime64_t tmp = cputime_to_cputime64(steal);
3987 struct rq *rq = this_rq();
3989 if (p == rq->idle) {
3990 p->stime = cputime_add(p->stime, steal);
3991 if (atomic_read(&rq->nr_iowait) > 0)
3992 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3994 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3996 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4000 * This function gets called by the timer code, with HZ frequency.
4001 * We call it with interrupts disabled.
4003 * It also gets called by the fork code, when changing the parent's
4006 void scheduler_tick(void)
4008 int cpu = smp_processor_id();
4009 struct rq *rq = cpu_rq(cpu);
4010 struct task_struct *curr = rq->curr;
4014 spin_lock(&rq->lock);
4015 update_rq_clock(rq);
4016 update_cpu_load(rq);
4017 curr->sched_class->task_tick(rq, curr, 0);
4018 spin_unlock(&rq->lock);
4021 rq->idle_at_tick = idle_cpu(cpu);
4022 trigger_load_balance(rq, cpu);
4026 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4028 void __kprobes add_preempt_count(int val)
4033 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4035 preempt_count() += val;
4037 * Spinlock count overflowing soon?
4039 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4042 EXPORT_SYMBOL(add_preempt_count);
4044 void __kprobes sub_preempt_count(int val)
4049 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4052 * Is the spinlock portion underflowing?
4054 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4055 !(preempt_count() & PREEMPT_MASK)))
4058 preempt_count() -= val;
4060 EXPORT_SYMBOL(sub_preempt_count);
4065 * Print scheduling while atomic bug:
4067 static noinline void __schedule_bug(struct task_struct *prev)
4069 struct pt_regs *regs = get_irq_regs();
4071 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4072 prev->comm, prev->pid, preempt_count());
4074 debug_show_held_locks(prev);
4076 if (irqs_disabled())
4077 print_irqtrace_events(prev);
4086 * Various schedule()-time debugging checks and statistics:
4088 static inline void schedule_debug(struct task_struct *prev)
4091 * Test if we are atomic. Since do_exit() needs to call into
4092 * schedule() atomically, we ignore that path for now.
4093 * Otherwise, whine if we are scheduling when we should not be.
4095 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4096 __schedule_bug(prev);
4098 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4100 schedstat_inc(this_rq(), sched_count);
4101 #ifdef CONFIG_SCHEDSTATS
4102 if (unlikely(prev->lock_depth >= 0)) {
4103 schedstat_inc(this_rq(), bkl_count);
4104 schedstat_inc(prev, sched_info.bkl_count);
4110 * Pick up the highest-prio task:
4112 static inline struct task_struct *
4113 pick_next_task(struct rq *rq, struct task_struct *prev)
4115 const struct sched_class *class;
4116 struct task_struct *p;
4119 * Optimization: we know that if all tasks are in
4120 * the fair class we can call that function directly:
4122 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4123 p = fair_sched_class.pick_next_task(rq);
4128 class = sched_class_highest;
4130 p = class->pick_next_task(rq);
4134 * Will never be NULL as the idle class always
4135 * returns a non-NULL p:
4137 class = class->next;
4142 * schedule() is the main scheduler function.
4144 asmlinkage void __sched schedule(void)
4146 struct task_struct *prev, *next;
4147 unsigned long *switch_count;
4149 int cpu, hrtick = sched_feat(HRTICK);
4153 cpu = smp_processor_id();
4157 switch_count = &prev->nivcsw;
4159 release_kernel_lock(prev);
4160 need_resched_nonpreemptible:
4162 schedule_debug(prev);
4168 * Do the rq-clock update outside the rq lock:
4170 local_irq_disable();
4171 update_rq_clock(rq);
4172 spin_lock(&rq->lock);
4173 clear_tsk_need_resched(prev);
4175 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4176 if (unlikely(signal_pending_state(prev->state, prev)))
4177 prev->state = TASK_RUNNING;
4179 deactivate_task(rq, prev, 1);
4180 switch_count = &prev->nvcsw;
4184 if (prev->sched_class->pre_schedule)
4185 prev->sched_class->pre_schedule(rq, prev);
4188 if (unlikely(!rq->nr_running))
4189 idle_balance(cpu, rq);
4191 prev->sched_class->put_prev_task(rq, prev);
4192 next = pick_next_task(rq, prev);
4194 if (likely(prev != next)) {
4195 sched_info_switch(prev, next);
4201 context_switch(rq, prev, next); /* unlocks the rq */
4203 * the context switch might have flipped the stack from under
4204 * us, hence refresh the local variables.
4206 cpu = smp_processor_id();
4209 spin_unlock_irq(&rq->lock);
4214 if (unlikely(reacquire_kernel_lock(current) < 0))
4215 goto need_resched_nonpreemptible;
4217 preempt_enable_no_resched();
4218 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4221 EXPORT_SYMBOL(schedule);
4223 #ifdef CONFIG_PREEMPT
4225 * this is the entry point to schedule() from in-kernel preemption
4226 * off of preempt_enable. Kernel preemptions off return from interrupt
4227 * occur there and call schedule directly.
4229 asmlinkage void __sched preempt_schedule(void)
4231 struct thread_info *ti = current_thread_info();
4234 * If there is a non-zero preempt_count or interrupts are disabled,
4235 * we do not want to preempt the current task. Just return..
4237 if (likely(ti->preempt_count || irqs_disabled()))
4241 add_preempt_count(PREEMPT_ACTIVE);
4243 sub_preempt_count(PREEMPT_ACTIVE);
4246 * Check again in case we missed a preemption opportunity
4247 * between schedule and now.
4250 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4252 EXPORT_SYMBOL(preempt_schedule);
4255 * this is the entry point to schedule() from kernel preemption
4256 * off of irq context.
4257 * Note, that this is called and return with irqs disabled. This will
4258 * protect us against recursive calling from irq.
4260 asmlinkage void __sched preempt_schedule_irq(void)
4262 struct thread_info *ti = current_thread_info();
4264 /* Catch callers which need to be fixed */
4265 BUG_ON(ti->preempt_count || !irqs_disabled());
4268 add_preempt_count(PREEMPT_ACTIVE);
4271 local_irq_disable();
4272 sub_preempt_count(PREEMPT_ACTIVE);
4275 * Check again in case we missed a preemption opportunity
4276 * between schedule and now.
4279 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4282 #endif /* CONFIG_PREEMPT */
4284 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4287 return try_to_wake_up(curr->private, mode, sync);
4289 EXPORT_SYMBOL(default_wake_function);
4292 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4293 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4294 * number) then we wake all the non-exclusive tasks and one exclusive task.
4296 * There are circumstances in which we can try to wake a task which has already
4297 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4298 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4300 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4301 int nr_exclusive, int sync, void *key)
4303 wait_queue_t *curr, *next;
4305 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4306 unsigned flags = curr->flags;
4308 if (curr->func(curr, mode, sync, key) &&
4309 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4315 * __wake_up - wake up threads blocked on a waitqueue.
4317 * @mode: which threads
4318 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4319 * @key: is directly passed to the wakeup function
4321 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4322 int nr_exclusive, void *key)
4324 unsigned long flags;
4326 spin_lock_irqsave(&q->lock, flags);
4327 __wake_up_common(q, mode, nr_exclusive, 0, key);
4328 spin_unlock_irqrestore(&q->lock, flags);
4330 EXPORT_SYMBOL(__wake_up);
4333 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4335 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4337 __wake_up_common(q, mode, 1, 0, NULL);
4341 * __wake_up_sync - wake up threads blocked on a waitqueue.
4343 * @mode: which threads
4344 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4346 * The sync wakeup differs that the waker knows that it will schedule
4347 * away soon, so while the target thread will be woken up, it will not
4348 * be migrated to another CPU - ie. the two threads are 'synchronized'
4349 * with each other. This can prevent needless bouncing between CPUs.
4351 * On UP it can prevent extra preemption.
4354 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4356 unsigned long flags;
4362 if (unlikely(!nr_exclusive))
4365 spin_lock_irqsave(&q->lock, flags);
4366 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4367 spin_unlock_irqrestore(&q->lock, flags);
4369 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4371 void complete(struct completion *x)
4373 unsigned long flags;
4375 spin_lock_irqsave(&x->wait.lock, flags);
4377 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4378 spin_unlock_irqrestore(&x->wait.lock, flags);
4380 EXPORT_SYMBOL(complete);
4382 void complete_all(struct completion *x)
4384 unsigned long flags;
4386 spin_lock_irqsave(&x->wait.lock, flags);
4387 x->done += UINT_MAX/2;
4388 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4389 spin_unlock_irqrestore(&x->wait.lock, flags);
4391 EXPORT_SYMBOL(complete_all);
4393 static inline long __sched
4394 do_wait_for_common(struct completion *x, long timeout, int state)
4397 DECLARE_WAITQUEUE(wait, current);
4399 wait.flags |= WQ_FLAG_EXCLUSIVE;
4400 __add_wait_queue_tail(&x->wait, &wait);
4402 if ((state == TASK_INTERRUPTIBLE &&
4403 signal_pending(current)) ||
4404 (state == TASK_KILLABLE &&
4405 fatal_signal_pending(current))) {
4406 timeout = -ERESTARTSYS;
4409 __set_current_state(state);
4410 spin_unlock_irq(&x->wait.lock);
4411 timeout = schedule_timeout(timeout);
4412 spin_lock_irq(&x->wait.lock);
4413 } while (!x->done && timeout);
4414 __remove_wait_queue(&x->wait, &wait);
4419 return timeout ?: 1;
4423 wait_for_common(struct completion *x, long timeout, int state)
4427 spin_lock_irq(&x->wait.lock);
4428 timeout = do_wait_for_common(x, timeout, state);
4429 spin_unlock_irq(&x->wait.lock);
4433 void __sched wait_for_completion(struct completion *x)
4435 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4437 EXPORT_SYMBOL(wait_for_completion);
4439 unsigned long __sched
4440 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4442 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4444 EXPORT_SYMBOL(wait_for_completion_timeout);
4446 int __sched wait_for_completion_interruptible(struct completion *x)
4448 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4449 if (t == -ERESTARTSYS)
4453 EXPORT_SYMBOL(wait_for_completion_interruptible);
4455 unsigned long __sched
4456 wait_for_completion_interruptible_timeout(struct completion *x,
4457 unsigned long timeout)
4459 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4461 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4463 int __sched wait_for_completion_killable(struct completion *x)
4465 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4466 if (t == -ERESTARTSYS)
4470 EXPORT_SYMBOL(wait_for_completion_killable);
4473 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4475 unsigned long flags;
4478 init_waitqueue_entry(&wait, current);
4480 __set_current_state(state);
4482 spin_lock_irqsave(&q->lock, flags);
4483 __add_wait_queue(q, &wait);
4484 spin_unlock(&q->lock);
4485 timeout = schedule_timeout(timeout);
4486 spin_lock_irq(&q->lock);
4487 __remove_wait_queue(q, &wait);
4488 spin_unlock_irqrestore(&q->lock, flags);
4493 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4495 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4497 EXPORT_SYMBOL(interruptible_sleep_on);
4500 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4502 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4504 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4506 void __sched sleep_on(wait_queue_head_t *q)
4508 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4510 EXPORT_SYMBOL(sleep_on);
4512 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4514 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4516 EXPORT_SYMBOL(sleep_on_timeout);
4518 #ifdef CONFIG_RT_MUTEXES
4521 * rt_mutex_setprio - set the current priority of a task
4523 * @prio: prio value (kernel-internal form)
4525 * This function changes the 'effective' priority of a task. It does
4526 * not touch ->normal_prio like __setscheduler().
4528 * Used by the rt_mutex code to implement priority inheritance logic.
4530 void rt_mutex_setprio(struct task_struct *p, int prio)
4532 unsigned long flags;
4533 int oldprio, on_rq, running;
4535 const struct sched_class *prev_class = p->sched_class;
4537 BUG_ON(prio < 0 || prio > MAX_PRIO);
4539 rq = task_rq_lock(p, &flags);
4540 update_rq_clock(rq);
4543 on_rq = p->se.on_rq;
4544 running = task_current(rq, p);
4546 dequeue_task(rq, p, 0);
4548 p->sched_class->put_prev_task(rq, p);
4551 p->sched_class = &rt_sched_class;
4553 p->sched_class = &fair_sched_class;
4558 p->sched_class->set_curr_task(rq);
4560 enqueue_task(rq, p, 0);
4562 check_class_changed(rq, p, prev_class, oldprio, running);
4564 task_rq_unlock(rq, &flags);
4569 void set_user_nice(struct task_struct *p, long nice)
4571 int old_prio, delta, on_rq;
4572 unsigned long flags;
4575 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4578 * We have to be careful, if called from sys_setpriority(),
4579 * the task might be in the middle of scheduling on another CPU.
4581 rq = task_rq_lock(p, &flags);
4582 update_rq_clock(rq);
4584 * The RT priorities are set via sched_setscheduler(), but we still
4585 * allow the 'normal' nice value to be set - but as expected
4586 * it wont have any effect on scheduling until the task is
4587 * SCHED_FIFO/SCHED_RR:
4589 if (task_has_rt_policy(p)) {
4590 p->static_prio = NICE_TO_PRIO(nice);
4593 on_rq = p->se.on_rq;
4595 dequeue_task(rq, p, 0);
4599 p->static_prio = NICE_TO_PRIO(nice);
4602 p->prio = effective_prio(p);
4603 delta = p->prio - old_prio;
4606 enqueue_task(rq, p, 0);
4609 * If the task increased its priority or is running and
4610 * lowered its priority, then reschedule its CPU:
4612 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4613 resched_task(rq->curr);
4616 task_rq_unlock(rq, &flags);
4618 EXPORT_SYMBOL(set_user_nice);
4621 * can_nice - check if a task can reduce its nice value
4625 int can_nice(const struct task_struct *p, const int nice)
4627 /* convert nice value [19,-20] to rlimit style value [1,40] */
4628 int nice_rlim = 20 - nice;
4630 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4631 capable(CAP_SYS_NICE));
4634 #ifdef __ARCH_WANT_SYS_NICE
4637 * sys_nice - change the priority of the current process.
4638 * @increment: priority increment
4640 * sys_setpriority is a more generic, but much slower function that
4641 * does similar things.
4643 asmlinkage long sys_nice(int increment)
4648 * Setpriority might change our priority at the same moment.
4649 * We don't have to worry. Conceptually one call occurs first
4650 * and we have a single winner.
4652 if (increment < -40)
4657 nice = PRIO_TO_NICE(current->static_prio) + increment;
4663 if (increment < 0 && !can_nice(current, nice))
4666 retval = security_task_setnice(current, nice);
4670 set_user_nice(current, nice);
4677 * task_prio - return the priority value of a given task.
4678 * @p: the task in question.
4680 * This is the priority value as seen by users in /proc.
4681 * RT tasks are offset by -200. Normal tasks are centered
4682 * around 0, value goes from -16 to +15.
4684 int task_prio(const struct task_struct *p)
4686 return p->prio - MAX_RT_PRIO;
4690 * task_nice - return the nice value of a given task.
4691 * @p: the task in question.
4693 int task_nice(const struct task_struct *p)
4695 return TASK_NICE(p);
4697 EXPORT_SYMBOL(task_nice);
4700 * idle_cpu - is a given cpu idle currently?
4701 * @cpu: the processor in question.
4703 int idle_cpu(int cpu)
4705 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4709 * idle_task - return the idle task for a given cpu.
4710 * @cpu: the processor in question.
4712 struct task_struct *idle_task(int cpu)
4714 return cpu_rq(cpu)->idle;
4718 * find_process_by_pid - find a process with a matching PID value.
4719 * @pid: the pid in question.
4721 static struct task_struct *find_process_by_pid(pid_t pid)
4723 return pid ? find_task_by_vpid(pid) : current;
4726 /* Actually do priority change: must hold rq lock. */
4728 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4730 BUG_ON(p->se.on_rq);
4733 switch (p->policy) {
4737 p->sched_class = &fair_sched_class;
4741 p->sched_class = &rt_sched_class;
4745 p->rt_priority = prio;
4746 p->normal_prio = normal_prio(p);
4747 /* we are holding p->pi_lock already */
4748 p->prio = rt_mutex_getprio(p);
4753 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4754 * @p: the task in question.
4755 * @policy: new policy.
4756 * @param: structure containing the new RT priority.
4758 * NOTE that the task may be already dead.
4760 int sched_setscheduler(struct task_struct *p, int policy,
4761 struct sched_param *param)
4763 int retval, oldprio, oldpolicy = -1, on_rq, running;
4764 unsigned long flags;
4765 const struct sched_class *prev_class = p->sched_class;
4768 /* may grab non-irq protected spin_locks */
4769 BUG_ON(in_interrupt());
4771 /* double check policy once rq lock held */
4773 policy = oldpolicy = p->policy;
4774 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4775 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4776 policy != SCHED_IDLE)
4779 * Valid priorities for SCHED_FIFO and SCHED_RR are
4780 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4781 * SCHED_BATCH and SCHED_IDLE is 0.
4783 if (param->sched_priority < 0 ||
4784 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4785 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4787 if (rt_policy(policy) != (param->sched_priority != 0))
4791 * Allow unprivileged RT tasks to decrease priority:
4793 if (!capable(CAP_SYS_NICE)) {
4794 if (rt_policy(policy)) {
4795 unsigned long rlim_rtprio;
4797 if (!lock_task_sighand(p, &flags))
4799 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4800 unlock_task_sighand(p, &flags);
4802 /* can't set/change the rt policy */
4803 if (policy != p->policy && !rlim_rtprio)
4806 /* can't increase priority */
4807 if (param->sched_priority > p->rt_priority &&
4808 param->sched_priority > rlim_rtprio)
4812 * Like positive nice levels, dont allow tasks to
4813 * move out of SCHED_IDLE either:
4815 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4818 /* can't change other user's priorities */
4819 if ((current->euid != p->euid) &&
4820 (current->euid != p->uid))
4824 #ifdef CONFIG_RT_GROUP_SCHED
4826 * Do not allow realtime tasks into groups that have no runtime
4829 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
4833 retval = security_task_setscheduler(p, policy, param);
4837 * make sure no PI-waiters arrive (or leave) while we are
4838 * changing the priority of the task:
4840 spin_lock_irqsave(&p->pi_lock, flags);
4842 * To be able to change p->policy safely, the apropriate
4843 * runqueue lock must be held.
4845 rq = __task_rq_lock(p);
4846 /* recheck policy now with rq lock held */
4847 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4848 policy = oldpolicy = -1;
4849 __task_rq_unlock(rq);
4850 spin_unlock_irqrestore(&p->pi_lock, flags);
4853 update_rq_clock(rq);
4854 on_rq = p->se.on_rq;
4855 running = task_current(rq, p);
4857 deactivate_task(rq, p, 0);
4859 p->sched_class->put_prev_task(rq, p);
4862 __setscheduler(rq, p, policy, param->sched_priority);
4865 p->sched_class->set_curr_task(rq);
4867 activate_task(rq, p, 0);
4869 check_class_changed(rq, p, prev_class, oldprio, running);
4871 __task_rq_unlock(rq);
4872 spin_unlock_irqrestore(&p->pi_lock, flags);
4874 rt_mutex_adjust_pi(p);
4878 EXPORT_SYMBOL_GPL(sched_setscheduler);
4881 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4883 struct sched_param lparam;
4884 struct task_struct *p;
4887 if (!param || pid < 0)
4889 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4894 p = find_process_by_pid(pid);
4896 retval = sched_setscheduler(p, policy, &lparam);
4903 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4904 * @pid: the pid in question.
4905 * @policy: new policy.
4906 * @param: structure containing the new RT priority.
4909 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4911 /* negative values for policy are not valid */
4915 return do_sched_setscheduler(pid, policy, param);
4919 * sys_sched_setparam - set/change the RT priority of a thread
4920 * @pid: the pid in question.
4921 * @param: structure containing the new RT priority.
4923 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4925 return do_sched_setscheduler(pid, -1, param);
4929 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4930 * @pid: the pid in question.
4932 asmlinkage long sys_sched_getscheduler(pid_t pid)
4934 struct task_struct *p;
4941 read_lock(&tasklist_lock);
4942 p = find_process_by_pid(pid);
4944 retval = security_task_getscheduler(p);
4948 read_unlock(&tasklist_lock);
4953 * sys_sched_getscheduler - get the RT priority of a thread
4954 * @pid: the pid in question.
4955 * @param: structure containing the RT priority.
4957 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4959 struct sched_param lp;
4960 struct task_struct *p;
4963 if (!param || pid < 0)
4966 read_lock(&tasklist_lock);
4967 p = find_process_by_pid(pid);
4972 retval = security_task_getscheduler(p);
4976 lp.sched_priority = p->rt_priority;
4977 read_unlock(&tasklist_lock);
4980 * This one might sleep, we cannot do it with a spinlock held ...
4982 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4987 read_unlock(&tasklist_lock);
4991 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
4993 cpumask_t cpus_allowed;
4994 cpumask_t new_mask = *in_mask;
4995 struct task_struct *p;
4999 read_lock(&tasklist_lock);
5001 p = find_process_by_pid(pid);
5003 read_unlock(&tasklist_lock);
5009 * It is not safe to call set_cpus_allowed with the
5010 * tasklist_lock held. We will bump the task_struct's
5011 * usage count and then drop tasklist_lock.
5014 read_unlock(&tasklist_lock);
5017 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5018 !capable(CAP_SYS_NICE))
5021 retval = security_task_setscheduler(p, 0, NULL);
5025 cpuset_cpus_allowed(p, &cpus_allowed);
5026 cpus_and(new_mask, new_mask, cpus_allowed);
5028 retval = set_cpus_allowed_ptr(p, &new_mask);
5031 cpuset_cpus_allowed(p, &cpus_allowed);
5032 if (!cpus_subset(new_mask, cpus_allowed)) {
5034 * We must have raced with a concurrent cpuset
5035 * update. Just reset the cpus_allowed to the
5036 * cpuset's cpus_allowed
5038 new_mask = cpus_allowed;
5048 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5049 cpumask_t *new_mask)
5051 if (len < sizeof(cpumask_t)) {
5052 memset(new_mask, 0, sizeof(cpumask_t));
5053 } else if (len > sizeof(cpumask_t)) {
5054 len = sizeof(cpumask_t);
5056 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5060 * sys_sched_setaffinity - set the cpu affinity of a process
5061 * @pid: pid of the process
5062 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5063 * @user_mask_ptr: user-space pointer to the new cpu mask
5065 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5066 unsigned long __user *user_mask_ptr)
5071 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5075 return sched_setaffinity(pid, &new_mask);
5078 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5080 struct task_struct *p;
5084 read_lock(&tasklist_lock);
5087 p = find_process_by_pid(pid);
5091 retval = security_task_getscheduler(p);
5095 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5098 read_unlock(&tasklist_lock);
5105 * sys_sched_getaffinity - get the cpu affinity of a process
5106 * @pid: pid of the process
5107 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5108 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5110 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5111 unsigned long __user *user_mask_ptr)
5116 if (len < sizeof(cpumask_t))
5119 ret = sched_getaffinity(pid, &mask);
5123 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5126 return sizeof(cpumask_t);
5130 * sys_sched_yield - yield the current processor to other threads.
5132 * This function yields the current CPU to other tasks. If there are no
5133 * other threads running on this CPU then this function will return.
5135 asmlinkage long sys_sched_yield(void)
5137 struct rq *rq = this_rq_lock();
5139 schedstat_inc(rq, yld_count);
5140 current->sched_class->yield_task(rq);
5143 * Since we are going to call schedule() anyway, there's
5144 * no need to preempt or enable interrupts:
5146 __release(rq->lock);
5147 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5148 _raw_spin_unlock(&rq->lock);
5149 preempt_enable_no_resched();
5156 static void __cond_resched(void)
5158 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5159 __might_sleep(__FILE__, __LINE__);
5162 * The BKS might be reacquired before we have dropped
5163 * PREEMPT_ACTIVE, which could trigger a second
5164 * cond_resched() call.
5167 add_preempt_count(PREEMPT_ACTIVE);
5169 sub_preempt_count(PREEMPT_ACTIVE);
5170 } while (need_resched());
5173 int __sched _cond_resched(void)
5175 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5176 system_state == SYSTEM_RUNNING) {
5182 EXPORT_SYMBOL(_cond_resched);
5185 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5186 * call schedule, and on return reacquire the lock.
5188 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5189 * operations here to prevent schedule() from being called twice (once via
5190 * spin_unlock(), once by hand).
5192 int cond_resched_lock(spinlock_t *lock)
5194 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5197 if (spin_needbreak(lock) || resched) {
5199 if (resched && need_resched())
5208 EXPORT_SYMBOL(cond_resched_lock);
5210 int __sched cond_resched_softirq(void)
5212 BUG_ON(!in_softirq());
5214 if (need_resched() && system_state == SYSTEM_RUNNING) {
5222 EXPORT_SYMBOL(cond_resched_softirq);
5225 * yield - yield the current processor to other threads.
5227 * This is a shortcut for kernel-space yielding - it marks the
5228 * thread runnable and calls sys_sched_yield().
5230 void __sched yield(void)
5232 set_current_state(TASK_RUNNING);
5235 EXPORT_SYMBOL(yield);
5238 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5239 * that process accounting knows that this is a task in IO wait state.
5241 * But don't do that if it is a deliberate, throttling IO wait (this task
5242 * has set its backing_dev_info: the queue against which it should throttle)
5244 void __sched io_schedule(void)
5246 struct rq *rq = &__raw_get_cpu_var(runqueues);
5248 delayacct_blkio_start();
5249 atomic_inc(&rq->nr_iowait);
5251 atomic_dec(&rq->nr_iowait);
5252 delayacct_blkio_end();
5254 EXPORT_SYMBOL(io_schedule);
5256 long __sched io_schedule_timeout(long timeout)
5258 struct rq *rq = &__raw_get_cpu_var(runqueues);
5261 delayacct_blkio_start();
5262 atomic_inc(&rq->nr_iowait);
5263 ret = schedule_timeout(timeout);
5264 atomic_dec(&rq->nr_iowait);
5265 delayacct_blkio_end();
5270 * sys_sched_get_priority_max - return maximum RT priority.
5271 * @policy: scheduling class.
5273 * this syscall returns the maximum rt_priority that can be used
5274 * by a given scheduling class.
5276 asmlinkage long sys_sched_get_priority_max(int policy)
5283 ret = MAX_USER_RT_PRIO-1;
5295 * sys_sched_get_priority_min - return minimum RT priority.
5296 * @policy: scheduling class.
5298 * this syscall returns the minimum rt_priority that can be used
5299 * by a given scheduling class.
5301 asmlinkage long sys_sched_get_priority_min(int policy)
5319 * sys_sched_rr_get_interval - return the default timeslice of a process.
5320 * @pid: pid of the process.
5321 * @interval: userspace pointer to the timeslice value.
5323 * this syscall writes the default timeslice value of a given process
5324 * into the user-space timespec buffer. A value of '0' means infinity.
5327 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5329 struct task_struct *p;
5330 unsigned int time_slice;
5338 read_lock(&tasklist_lock);
5339 p = find_process_by_pid(pid);
5343 retval = security_task_getscheduler(p);
5348 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5349 * tasks that are on an otherwise idle runqueue:
5352 if (p->policy == SCHED_RR) {
5353 time_slice = DEF_TIMESLICE;
5354 } else if (p->policy != SCHED_FIFO) {
5355 struct sched_entity *se = &p->se;
5356 unsigned long flags;
5359 rq = task_rq_lock(p, &flags);
5360 if (rq->cfs.load.weight)
5361 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5362 task_rq_unlock(rq, &flags);
5364 read_unlock(&tasklist_lock);
5365 jiffies_to_timespec(time_slice, &t);
5366 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5370 read_unlock(&tasklist_lock);
5374 static const char stat_nam[] = "RSDTtZX";
5376 void sched_show_task(struct task_struct *p)
5378 unsigned long free = 0;
5381 state = p->state ? __ffs(p->state) + 1 : 0;
5382 printk(KERN_INFO "%-13.13s %c", p->comm,
5383 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5384 #if BITS_PER_LONG == 32
5385 if (state == TASK_RUNNING)
5386 printk(KERN_CONT " running ");
5388 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5390 if (state == TASK_RUNNING)
5391 printk(KERN_CONT " running task ");
5393 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5395 #ifdef CONFIG_DEBUG_STACK_USAGE
5397 unsigned long *n = end_of_stack(p);
5400 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5403 printk(KERN_CONT "%5lu %5d %6d\n", free,
5404 task_pid_nr(p), task_pid_nr(p->real_parent));
5406 show_stack(p, NULL);
5409 void show_state_filter(unsigned long state_filter)
5411 struct task_struct *g, *p;
5413 #if BITS_PER_LONG == 32
5415 " task PC stack pid father\n");
5418 " task PC stack pid father\n");
5420 read_lock(&tasklist_lock);
5421 do_each_thread(g, p) {
5423 * reset the NMI-timeout, listing all files on a slow
5424 * console might take alot of time:
5426 touch_nmi_watchdog();
5427 if (!state_filter || (p->state & state_filter))
5429 } while_each_thread(g, p);
5431 touch_all_softlockup_watchdogs();
5433 #ifdef CONFIG_SCHED_DEBUG
5434 sysrq_sched_debug_show();
5436 read_unlock(&tasklist_lock);
5438 * Only show locks if all tasks are dumped:
5440 if (state_filter == -1)
5441 debug_show_all_locks();
5444 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5446 idle->sched_class = &idle_sched_class;
5450 * init_idle - set up an idle thread for a given CPU
5451 * @idle: task in question
5452 * @cpu: cpu the idle task belongs to
5454 * NOTE: this function does not set the idle thread's NEED_RESCHED
5455 * flag, to make booting more robust.
5457 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5459 struct rq *rq = cpu_rq(cpu);
5460 unsigned long flags;
5463 idle->se.exec_start = sched_clock();
5465 idle->prio = idle->normal_prio = MAX_PRIO;
5466 idle->cpus_allowed = cpumask_of_cpu(cpu);
5467 __set_task_cpu(idle, cpu);
5469 spin_lock_irqsave(&rq->lock, flags);
5470 rq->curr = rq->idle = idle;
5471 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5474 spin_unlock_irqrestore(&rq->lock, flags);
5476 /* Set the preempt count _outside_ the spinlocks! */
5477 #if defined(CONFIG_PREEMPT)
5478 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5480 task_thread_info(idle)->preempt_count = 0;
5483 * The idle tasks have their own, simple scheduling class:
5485 idle->sched_class = &idle_sched_class;
5489 * In a system that switches off the HZ timer nohz_cpu_mask
5490 * indicates which cpus entered this state. This is used
5491 * in the rcu update to wait only for active cpus. For system
5492 * which do not switch off the HZ timer nohz_cpu_mask should
5493 * always be CPU_MASK_NONE.
5495 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5498 * Increase the granularity value when there are more CPUs,
5499 * because with more CPUs the 'effective latency' as visible
5500 * to users decreases. But the relationship is not linear,
5501 * so pick a second-best guess by going with the log2 of the
5504 * This idea comes from the SD scheduler of Con Kolivas:
5506 static inline void sched_init_granularity(void)
5508 unsigned int factor = 1 + ilog2(num_online_cpus());
5509 const unsigned long limit = 200000000;
5511 sysctl_sched_min_granularity *= factor;
5512 if (sysctl_sched_min_granularity > limit)
5513 sysctl_sched_min_granularity = limit;
5515 sysctl_sched_latency *= factor;
5516 if (sysctl_sched_latency > limit)
5517 sysctl_sched_latency = limit;
5519 sysctl_sched_wakeup_granularity *= factor;
5524 * This is how migration works:
5526 * 1) we queue a struct migration_req structure in the source CPU's
5527 * runqueue and wake up that CPU's migration thread.
5528 * 2) we down() the locked semaphore => thread blocks.
5529 * 3) migration thread wakes up (implicitly it forces the migrated
5530 * thread off the CPU)
5531 * 4) it gets the migration request and checks whether the migrated
5532 * task is still in the wrong runqueue.
5533 * 5) if it's in the wrong runqueue then the migration thread removes
5534 * it and puts it into the right queue.
5535 * 6) migration thread up()s the semaphore.
5536 * 7) we wake up and the migration is done.
5540 * Change a given task's CPU affinity. Migrate the thread to a
5541 * proper CPU and schedule it away if the CPU it's executing on
5542 * is removed from the allowed bitmask.
5544 * NOTE: the caller must have a valid reference to the task, the
5545 * task must not exit() & deallocate itself prematurely. The
5546 * call is not atomic; no spinlocks may be held.
5548 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5550 struct migration_req req;
5551 unsigned long flags;
5555 rq = task_rq_lock(p, &flags);
5556 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5561 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5562 !cpus_equal(p->cpus_allowed, *new_mask))) {
5567 if (p->sched_class->set_cpus_allowed)
5568 p->sched_class->set_cpus_allowed(p, new_mask);
5570 p->cpus_allowed = *new_mask;
5571 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5574 /* Can the task run on the task's current CPU? If so, we're done */
5575 if (cpu_isset(task_cpu(p), *new_mask))
5578 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5579 /* Need help from migration thread: drop lock and wait. */
5580 task_rq_unlock(rq, &flags);
5581 wake_up_process(rq->migration_thread);
5582 wait_for_completion(&req.done);
5583 tlb_migrate_finish(p->mm);
5587 task_rq_unlock(rq, &flags);
5591 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5594 * Move (not current) task off this cpu, onto dest cpu. We're doing
5595 * this because either it can't run here any more (set_cpus_allowed()
5596 * away from this CPU, or CPU going down), or because we're
5597 * attempting to rebalance this task on exec (sched_exec).
5599 * So we race with normal scheduler movements, but that's OK, as long
5600 * as the task is no longer on this CPU.
5602 * Returns non-zero if task was successfully migrated.
5604 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5606 struct rq *rq_dest, *rq_src;
5609 if (unlikely(cpu_is_offline(dest_cpu)))
5612 rq_src = cpu_rq(src_cpu);
5613 rq_dest = cpu_rq(dest_cpu);
5615 double_rq_lock(rq_src, rq_dest);
5616 /* Already moved. */
5617 if (task_cpu(p) != src_cpu)
5619 /* Affinity changed (again). */
5620 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5623 on_rq = p->se.on_rq;
5625 deactivate_task(rq_src, p, 0);
5627 set_task_cpu(p, dest_cpu);
5629 activate_task(rq_dest, p, 0);
5630 check_preempt_curr(rq_dest, p);
5634 double_rq_unlock(rq_src, rq_dest);
5639 * migration_thread - this is a highprio system thread that performs
5640 * thread migration by bumping thread off CPU then 'pushing' onto
5643 static int migration_thread(void *data)
5645 int cpu = (long)data;
5649 BUG_ON(rq->migration_thread != current);
5651 set_current_state(TASK_INTERRUPTIBLE);
5652 while (!kthread_should_stop()) {
5653 struct migration_req *req;
5654 struct list_head *head;
5656 spin_lock_irq(&rq->lock);
5658 if (cpu_is_offline(cpu)) {
5659 spin_unlock_irq(&rq->lock);
5663 if (rq->active_balance) {
5664 active_load_balance(rq, cpu);
5665 rq->active_balance = 0;
5668 head = &rq->migration_queue;
5670 if (list_empty(head)) {
5671 spin_unlock_irq(&rq->lock);
5673 set_current_state(TASK_INTERRUPTIBLE);
5676 req = list_entry(head->next, struct migration_req, list);
5677 list_del_init(head->next);
5679 spin_unlock(&rq->lock);
5680 __migrate_task(req->task, cpu, req->dest_cpu);
5683 complete(&req->done);
5685 __set_current_state(TASK_RUNNING);
5689 /* Wait for kthread_stop */
5690 set_current_state(TASK_INTERRUPTIBLE);
5691 while (!kthread_should_stop()) {
5693 set_current_state(TASK_INTERRUPTIBLE);
5695 __set_current_state(TASK_RUNNING);
5699 #ifdef CONFIG_HOTPLUG_CPU
5701 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5705 local_irq_disable();
5706 ret = __migrate_task(p, src_cpu, dest_cpu);
5712 * Figure out where task on dead CPU should go, use force if necessary.
5713 * NOTE: interrupts should be disabled by the caller
5715 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5717 unsigned long flags;
5724 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5725 cpus_and(mask, mask, p->cpus_allowed);
5726 dest_cpu = any_online_cpu(mask);
5728 /* On any allowed CPU? */
5729 if (dest_cpu >= nr_cpu_ids)
5730 dest_cpu = any_online_cpu(p->cpus_allowed);
5732 /* No more Mr. Nice Guy. */
5733 if (dest_cpu >= nr_cpu_ids) {
5734 cpumask_t cpus_allowed;
5736 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5738 * Try to stay on the same cpuset, where the
5739 * current cpuset may be a subset of all cpus.
5740 * The cpuset_cpus_allowed_locked() variant of
5741 * cpuset_cpus_allowed() will not block. It must be
5742 * called within calls to cpuset_lock/cpuset_unlock.
5744 rq = task_rq_lock(p, &flags);
5745 p->cpus_allowed = cpus_allowed;
5746 dest_cpu = any_online_cpu(p->cpus_allowed);
5747 task_rq_unlock(rq, &flags);
5750 * Don't tell them about moving exiting tasks or
5751 * kernel threads (both mm NULL), since they never
5754 if (p->mm && printk_ratelimit()) {
5755 printk(KERN_INFO "process %d (%s) no "
5756 "longer affine to cpu%d\n",
5757 task_pid_nr(p), p->comm, dead_cpu);
5760 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5764 * While a dead CPU has no uninterruptible tasks queued at this point,
5765 * it might still have a nonzero ->nr_uninterruptible counter, because
5766 * for performance reasons the counter is not stricly tracking tasks to
5767 * their home CPUs. So we just add the counter to another CPU's counter,
5768 * to keep the global sum constant after CPU-down:
5770 static void migrate_nr_uninterruptible(struct rq *rq_src)
5772 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5773 unsigned long flags;
5775 local_irq_save(flags);
5776 double_rq_lock(rq_src, rq_dest);
5777 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5778 rq_src->nr_uninterruptible = 0;
5779 double_rq_unlock(rq_src, rq_dest);
5780 local_irq_restore(flags);
5783 /* Run through task list and migrate tasks from the dead cpu. */
5784 static void migrate_live_tasks(int src_cpu)
5786 struct task_struct *p, *t;
5788 read_lock(&tasklist_lock);
5790 do_each_thread(t, p) {
5794 if (task_cpu(p) == src_cpu)
5795 move_task_off_dead_cpu(src_cpu, p);
5796 } while_each_thread(t, p);
5798 read_unlock(&tasklist_lock);
5802 * Schedules idle task to be the next runnable task on current CPU.
5803 * It does so by boosting its priority to highest possible.
5804 * Used by CPU offline code.
5806 void sched_idle_next(void)
5808 int this_cpu = smp_processor_id();
5809 struct rq *rq = cpu_rq(this_cpu);
5810 struct task_struct *p = rq->idle;
5811 unsigned long flags;
5813 /* cpu has to be offline */
5814 BUG_ON(cpu_online(this_cpu));
5817 * Strictly not necessary since rest of the CPUs are stopped by now
5818 * and interrupts disabled on the current cpu.
5820 spin_lock_irqsave(&rq->lock, flags);
5822 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5824 update_rq_clock(rq);
5825 activate_task(rq, p, 0);
5827 spin_unlock_irqrestore(&rq->lock, flags);
5831 * Ensures that the idle task is using init_mm right before its cpu goes
5834 void idle_task_exit(void)
5836 struct mm_struct *mm = current->active_mm;
5838 BUG_ON(cpu_online(smp_processor_id()));
5841 switch_mm(mm, &init_mm, current);
5845 /* called under rq->lock with disabled interrupts */
5846 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5848 struct rq *rq = cpu_rq(dead_cpu);
5850 /* Must be exiting, otherwise would be on tasklist. */
5851 BUG_ON(!p->exit_state);
5853 /* Cannot have done final schedule yet: would have vanished. */
5854 BUG_ON(p->state == TASK_DEAD);
5859 * Drop lock around migration; if someone else moves it,
5860 * that's OK. No task can be added to this CPU, so iteration is
5863 spin_unlock_irq(&rq->lock);
5864 move_task_off_dead_cpu(dead_cpu, p);
5865 spin_lock_irq(&rq->lock);
5870 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5871 static void migrate_dead_tasks(unsigned int dead_cpu)
5873 struct rq *rq = cpu_rq(dead_cpu);
5874 struct task_struct *next;
5877 if (!rq->nr_running)
5879 update_rq_clock(rq);
5880 next = pick_next_task(rq, rq->curr);
5883 migrate_dead(dead_cpu, next);
5887 #endif /* CONFIG_HOTPLUG_CPU */
5889 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5891 static struct ctl_table sd_ctl_dir[] = {
5893 .procname = "sched_domain",
5899 static struct ctl_table sd_ctl_root[] = {
5901 .ctl_name = CTL_KERN,
5902 .procname = "kernel",
5904 .child = sd_ctl_dir,
5909 static struct ctl_table *sd_alloc_ctl_entry(int n)
5911 struct ctl_table *entry =
5912 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5917 static void sd_free_ctl_entry(struct ctl_table **tablep)
5919 struct ctl_table *entry;
5922 * In the intermediate directories, both the child directory and
5923 * procname are dynamically allocated and could fail but the mode
5924 * will always be set. In the lowest directory the names are
5925 * static strings and all have proc handlers.
5927 for (entry = *tablep; entry->mode; entry++) {
5929 sd_free_ctl_entry(&entry->child);
5930 if (entry->proc_handler == NULL)
5931 kfree(entry->procname);
5939 set_table_entry(struct ctl_table *entry,
5940 const char *procname, void *data, int maxlen,
5941 mode_t mode, proc_handler *proc_handler)
5943 entry->procname = procname;
5945 entry->maxlen = maxlen;
5947 entry->proc_handler = proc_handler;
5950 static struct ctl_table *
5951 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5953 struct ctl_table *table = sd_alloc_ctl_entry(12);
5958 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5959 sizeof(long), 0644, proc_doulongvec_minmax);
5960 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5961 sizeof(long), 0644, proc_doulongvec_minmax);
5962 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5963 sizeof(int), 0644, proc_dointvec_minmax);
5964 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5965 sizeof(int), 0644, proc_dointvec_minmax);
5966 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5967 sizeof(int), 0644, proc_dointvec_minmax);
5968 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5969 sizeof(int), 0644, proc_dointvec_minmax);
5970 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5971 sizeof(int), 0644, proc_dointvec_minmax);
5972 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5973 sizeof(int), 0644, proc_dointvec_minmax);
5974 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5975 sizeof(int), 0644, proc_dointvec_minmax);
5976 set_table_entry(&table[9], "cache_nice_tries",
5977 &sd->cache_nice_tries,
5978 sizeof(int), 0644, proc_dointvec_minmax);
5979 set_table_entry(&table[10], "flags", &sd->flags,
5980 sizeof(int), 0644, proc_dointvec_minmax);
5981 /* &table[11] is terminator */
5986 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5988 struct ctl_table *entry, *table;
5989 struct sched_domain *sd;
5990 int domain_num = 0, i;
5993 for_each_domain(cpu, sd)
5995 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6000 for_each_domain(cpu, sd) {
6001 snprintf(buf, 32, "domain%d", i);
6002 entry->procname = kstrdup(buf, GFP_KERNEL);
6004 entry->child = sd_alloc_ctl_domain_table(sd);
6011 static struct ctl_table_header *sd_sysctl_header;
6012 static void register_sched_domain_sysctl(void)
6014 int i, cpu_num = num_online_cpus();
6015 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6018 WARN_ON(sd_ctl_dir[0].child);
6019 sd_ctl_dir[0].child = entry;
6024 for_each_online_cpu(i) {
6025 snprintf(buf, 32, "cpu%d", i);
6026 entry->procname = kstrdup(buf, GFP_KERNEL);
6028 entry->child = sd_alloc_ctl_cpu_table(i);
6032 WARN_ON(sd_sysctl_header);
6033 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6036 /* may be called multiple times per register */
6037 static void unregister_sched_domain_sysctl(void)
6039 if (sd_sysctl_header)
6040 unregister_sysctl_table(sd_sysctl_header);
6041 sd_sysctl_header = NULL;
6042 if (sd_ctl_dir[0].child)
6043 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6046 static void register_sched_domain_sysctl(void)
6049 static void unregister_sched_domain_sysctl(void)
6054 static void set_rq_online(struct rq *rq)
6057 const struct sched_class *class;
6059 cpu_set(rq->cpu, rq->rd->online);
6062 for_each_class(class) {
6063 if (class->rq_online)
6064 class->rq_online(rq);
6069 static void set_rq_offline(struct rq *rq)
6072 const struct sched_class *class;
6074 for_each_class(class) {
6075 if (class->rq_offline)
6076 class->rq_offline(rq);
6079 cpu_clear(rq->cpu, rq->rd->online);
6085 * migration_call - callback that gets triggered when a CPU is added.
6086 * Here we can start up the necessary migration thread for the new CPU.
6088 static int __cpuinit
6089 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6091 struct task_struct *p;
6092 int cpu = (long)hcpu;
6093 unsigned long flags;
6098 case CPU_UP_PREPARE:
6099 case CPU_UP_PREPARE_FROZEN:
6100 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6103 kthread_bind(p, cpu);
6104 /* Must be high prio: stop_machine expects to yield to it. */
6105 rq = task_rq_lock(p, &flags);
6106 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6107 task_rq_unlock(rq, &flags);
6108 cpu_rq(cpu)->migration_thread = p;
6112 case CPU_ONLINE_FROZEN:
6113 /* Strictly unnecessary, as first user will wake it. */
6114 wake_up_process(cpu_rq(cpu)->migration_thread);
6116 /* Update our root-domain */
6118 spin_lock_irqsave(&rq->lock, flags);
6120 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6124 spin_unlock_irqrestore(&rq->lock, flags);
6127 #ifdef CONFIG_HOTPLUG_CPU
6128 case CPU_UP_CANCELED:
6129 case CPU_UP_CANCELED_FROZEN:
6130 if (!cpu_rq(cpu)->migration_thread)
6132 /* Unbind it from offline cpu so it can run. Fall thru. */
6133 kthread_bind(cpu_rq(cpu)->migration_thread,
6134 any_online_cpu(cpu_online_map));
6135 kthread_stop(cpu_rq(cpu)->migration_thread);
6136 cpu_rq(cpu)->migration_thread = NULL;
6140 case CPU_DEAD_FROZEN:
6141 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6142 migrate_live_tasks(cpu);
6144 kthread_stop(rq->migration_thread);
6145 rq->migration_thread = NULL;
6146 /* Idle task back to normal (off runqueue, low prio) */
6147 spin_lock_irq(&rq->lock);
6148 update_rq_clock(rq);
6149 deactivate_task(rq, rq->idle, 0);
6150 rq->idle->static_prio = MAX_PRIO;
6151 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6152 rq->idle->sched_class = &idle_sched_class;
6153 migrate_dead_tasks(cpu);
6154 spin_unlock_irq(&rq->lock);
6156 migrate_nr_uninterruptible(rq);
6157 BUG_ON(rq->nr_running != 0);
6160 * No need to migrate the tasks: it was best-effort if
6161 * they didn't take sched_hotcpu_mutex. Just wake up
6164 spin_lock_irq(&rq->lock);
6165 while (!list_empty(&rq->migration_queue)) {
6166 struct migration_req *req;
6168 req = list_entry(rq->migration_queue.next,
6169 struct migration_req, list);
6170 list_del_init(&req->list);
6171 complete(&req->done);
6173 spin_unlock_irq(&rq->lock);
6177 case CPU_DYING_FROZEN:
6178 /* Update our root-domain */
6180 spin_lock_irqsave(&rq->lock, flags);
6182 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6185 spin_unlock_irqrestore(&rq->lock, flags);
6192 /* Register at highest priority so that task migration (migrate_all_tasks)
6193 * happens before everything else.
6195 static struct notifier_block __cpuinitdata migration_notifier = {
6196 .notifier_call = migration_call,
6200 void __init migration_init(void)
6202 void *cpu = (void *)(long)smp_processor_id();
6205 /* Start one for the boot CPU: */
6206 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6207 BUG_ON(err == NOTIFY_BAD);
6208 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6209 register_cpu_notifier(&migration_notifier);
6215 #ifdef CONFIG_SCHED_DEBUG
6217 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6230 case SD_LV_ALLNODES:
6239 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6240 cpumask_t *groupmask)
6242 struct sched_group *group = sd->groups;
6245 cpulist_scnprintf(str, sizeof(str), sd->span);
6246 cpus_clear(*groupmask);
6248 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6250 if (!(sd->flags & SD_LOAD_BALANCE)) {
6251 printk("does not load-balance\n");
6253 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6258 printk(KERN_CONT "span %s level %s\n",
6259 str, sd_level_to_string(sd->level));
6261 if (!cpu_isset(cpu, sd->span)) {
6262 printk(KERN_ERR "ERROR: domain->span does not contain "
6265 if (!cpu_isset(cpu, group->cpumask)) {
6266 printk(KERN_ERR "ERROR: domain->groups does not contain"
6270 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6274 printk(KERN_ERR "ERROR: group is NULL\n");
6278 if (!group->__cpu_power) {
6279 printk(KERN_CONT "\n");
6280 printk(KERN_ERR "ERROR: domain->cpu_power not "
6285 if (!cpus_weight(group->cpumask)) {
6286 printk(KERN_CONT "\n");
6287 printk(KERN_ERR "ERROR: empty group\n");
6291 if (cpus_intersects(*groupmask, group->cpumask)) {
6292 printk(KERN_CONT "\n");
6293 printk(KERN_ERR "ERROR: repeated CPUs\n");
6297 cpus_or(*groupmask, *groupmask, group->cpumask);
6299 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6300 printk(KERN_CONT " %s", str);
6302 group = group->next;
6303 } while (group != sd->groups);
6304 printk(KERN_CONT "\n");
6306 if (!cpus_equal(sd->span, *groupmask))
6307 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6309 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6310 printk(KERN_ERR "ERROR: parent span is not a superset "
6311 "of domain->span\n");
6315 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6317 cpumask_t *groupmask;
6321 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6325 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6327 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6329 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6334 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6343 #else /* !CONFIG_SCHED_DEBUG */
6344 # define sched_domain_debug(sd, cpu) do { } while (0)
6345 #endif /* CONFIG_SCHED_DEBUG */
6347 static int sd_degenerate(struct sched_domain *sd)
6349 if (cpus_weight(sd->span) == 1)
6352 /* Following flags need at least 2 groups */
6353 if (sd->flags & (SD_LOAD_BALANCE |
6354 SD_BALANCE_NEWIDLE |
6358 SD_SHARE_PKG_RESOURCES)) {
6359 if (sd->groups != sd->groups->next)
6363 /* Following flags don't use groups */
6364 if (sd->flags & (SD_WAKE_IDLE |
6373 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6375 unsigned long cflags = sd->flags, pflags = parent->flags;
6377 if (sd_degenerate(parent))
6380 if (!cpus_equal(sd->span, parent->span))
6383 /* Does parent contain flags not in child? */
6384 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6385 if (cflags & SD_WAKE_AFFINE)
6386 pflags &= ~SD_WAKE_BALANCE;
6387 /* Flags needing groups don't count if only 1 group in parent */
6388 if (parent->groups == parent->groups->next) {
6389 pflags &= ~(SD_LOAD_BALANCE |
6390 SD_BALANCE_NEWIDLE |
6394 SD_SHARE_PKG_RESOURCES);
6396 if (~cflags & pflags)
6402 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6404 unsigned long flags;
6406 spin_lock_irqsave(&rq->lock, flags);
6409 struct root_domain *old_rd = rq->rd;
6411 if (cpu_isset(rq->cpu, old_rd->online))
6414 cpu_clear(rq->cpu, old_rd->span);
6416 if (atomic_dec_and_test(&old_rd->refcount))
6420 atomic_inc(&rd->refcount);
6423 cpu_set(rq->cpu, rd->span);
6424 if (cpu_isset(rq->cpu, cpu_online_map))
6427 spin_unlock_irqrestore(&rq->lock, flags);
6430 static void init_rootdomain(struct root_domain *rd)
6432 memset(rd, 0, sizeof(*rd));
6434 cpus_clear(rd->span);
6435 cpus_clear(rd->online);
6437 cpupri_init(&rd->cpupri);
6440 static void init_defrootdomain(void)
6442 init_rootdomain(&def_root_domain);
6443 atomic_set(&def_root_domain.refcount, 1);
6446 static struct root_domain *alloc_rootdomain(void)
6448 struct root_domain *rd;
6450 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6454 init_rootdomain(rd);
6460 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6461 * hold the hotplug lock.
6464 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6466 struct rq *rq = cpu_rq(cpu);
6467 struct sched_domain *tmp;
6469 /* Remove the sched domains which do not contribute to scheduling. */
6470 for (tmp = sd; tmp; tmp = tmp->parent) {
6471 struct sched_domain *parent = tmp->parent;
6474 if (sd_parent_degenerate(tmp, parent)) {
6475 tmp->parent = parent->parent;
6477 parent->parent->child = tmp;
6481 if (sd && sd_degenerate(sd)) {
6487 sched_domain_debug(sd, cpu);
6489 rq_attach_root(rq, rd);
6490 rcu_assign_pointer(rq->sd, sd);
6493 /* cpus with isolated domains */
6494 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6496 /* Setup the mask of cpus configured for isolated domains */
6497 static int __init isolated_cpu_setup(char *str)
6499 int ints[NR_CPUS], i;
6501 str = get_options(str, ARRAY_SIZE(ints), ints);
6502 cpus_clear(cpu_isolated_map);
6503 for (i = 1; i <= ints[0]; i++)
6504 if (ints[i] < NR_CPUS)
6505 cpu_set(ints[i], cpu_isolated_map);
6509 __setup("isolcpus=", isolated_cpu_setup);
6512 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6513 * to a function which identifies what group(along with sched group) a CPU
6514 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6515 * (due to the fact that we keep track of groups covered with a cpumask_t).
6517 * init_sched_build_groups will build a circular linked list of the groups
6518 * covered by the given span, and will set each group's ->cpumask correctly,
6519 * and ->cpu_power to 0.
6522 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6523 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6524 struct sched_group **sg,
6525 cpumask_t *tmpmask),
6526 cpumask_t *covered, cpumask_t *tmpmask)
6528 struct sched_group *first = NULL, *last = NULL;
6531 cpus_clear(*covered);
6533 for_each_cpu_mask(i, *span) {
6534 struct sched_group *sg;
6535 int group = group_fn(i, cpu_map, &sg, tmpmask);
6538 if (cpu_isset(i, *covered))
6541 cpus_clear(sg->cpumask);
6542 sg->__cpu_power = 0;
6544 for_each_cpu_mask(j, *span) {
6545 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6548 cpu_set(j, *covered);
6549 cpu_set(j, sg->cpumask);
6560 #define SD_NODES_PER_DOMAIN 16
6565 * find_next_best_node - find the next node to include in a sched_domain
6566 * @node: node whose sched_domain we're building
6567 * @used_nodes: nodes already in the sched_domain
6569 * Find the next node to include in a given scheduling domain. Simply
6570 * finds the closest node not already in the @used_nodes map.
6572 * Should use nodemask_t.
6574 static int find_next_best_node(int node, nodemask_t *used_nodes)
6576 int i, n, val, min_val, best_node = 0;
6580 for (i = 0; i < MAX_NUMNODES; i++) {
6581 /* Start at @node */
6582 n = (node + i) % MAX_NUMNODES;
6584 if (!nr_cpus_node(n))
6587 /* Skip already used nodes */
6588 if (node_isset(n, *used_nodes))
6591 /* Simple min distance search */
6592 val = node_distance(node, n);
6594 if (val < min_val) {
6600 node_set(best_node, *used_nodes);
6605 * sched_domain_node_span - get a cpumask for a node's sched_domain
6606 * @node: node whose cpumask we're constructing
6607 * @span: resulting cpumask
6609 * Given a node, construct a good cpumask for its sched_domain to span. It
6610 * should be one that prevents unnecessary balancing, but also spreads tasks
6613 static void sched_domain_node_span(int node, cpumask_t *span)
6615 nodemask_t used_nodes;
6616 node_to_cpumask_ptr(nodemask, node);
6620 nodes_clear(used_nodes);
6622 cpus_or(*span, *span, *nodemask);
6623 node_set(node, used_nodes);
6625 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6626 int next_node = find_next_best_node(node, &used_nodes);
6628 node_to_cpumask_ptr_next(nodemask, next_node);
6629 cpus_or(*span, *span, *nodemask);
6632 #endif /* CONFIG_NUMA */
6634 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6637 * SMT sched-domains:
6639 #ifdef CONFIG_SCHED_SMT
6640 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6641 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6644 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6648 *sg = &per_cpu(sched_group_cpus, cpu);
6651 #endif /* CONFIG_SCHED_SMT */
6654 * multi-core sched-domains:
6656 #ifdef CONFIG_SCHED_MC
6657 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6658 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6659 #endif /* CONFIG_SCHED_MC */
6661 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6663 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6668 *mask = per_cpu(cpu_sibling_map, cpu);
6669 cpus_and(*mask, *mask, *cpu_map);
6670 group = first_cpu(*mask);
6672 *sg = &per_cpu(sched_group_core, group);
6675 #elif defined(CONFIG_SCHED_MC)
6677 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6681 *sg = &per_cpu(sched_group_core, cpu);
6686 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6687 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6690 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6694 #ifdef CONFIG_SCHED_MC
6695 *mask = cpu_coregroup_map(cpu);
6696 cpus_and(*mask, *mask, *cpu_map);
6697 group = first_cpu(*mask);
6698 #elif defined(CONFIG_SCHED_SMT)
6699 *mask = per_cpu(cpu_sibling_map, cpu);
6700 cpus_and(*mask, *mask, *cpu_map);
6701 group = first_cpu(*mask);
6706 *sg = &per_cpu(sched_group_phys, group);
6712 * The init_sched_build_groups can't handle what we want to do with node
6713 * groups, so roll our own. Now each node has its own list of groups which
6714 * gets dynamically allocated.
6716 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6717 static struct sched_group ***sched_group_nodes_bycpu;
6719 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6720 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6722 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6723 struct sched_group **sg, cpumask_t *nodemask)
6727 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6728 cpus_and(*nodemask, *nodemask, *cpu_map);
6729 group = first_cpu(*nodemask);
6732 *sg = &per_cpu(sched_group_allnodes, group);
6736 static void init_numa_sched_groups_power(struct sched_group *group_head)
6738 struct sched_group *sg = group_head;
6744 for_each_cpu_mask(j, sg->cpumask) {
6745 struct sched_domain *sd;
6747 sd = &per_cpu(phys_domains, j);
6748 if (j != first_cpu(sd->groups->cpumask)) {
6750 * Only add "power" once for each
6756 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6759 } while (sg != group_head);
6761 #endif /* CONFIG_NUMA */
6764 /* Free memory allocated for various sched_group structures */
6765 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6769 for_each_cpu_mask(cpu, *cpu_map) {
6770 struct sched_group **sched_group_nodes
6771 = sched_group_nodes_bycpu[cpu];
6773 if (!sched_group_nodes)
6776 for (i = 0; i < MAX_NUMNODES; i++) {
6777 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6779 *nodemask = node_to_cpumask(i);
6780 cpus_and(*nodemask, *nodemask, *cpu_map);
6781 if (cpus_empty(*nodemask))
6791 if (oldsg != sched_group_nodes[i])
6794 kfree(sched_group_nodes);
6795 sched_group_nodes_bycpu[cpu] = NULL;
6798 #else /* !CONFIG_NUMA */
6799 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6802 #endif /* CONFIG_NUMA */
6805 * Initialize sched groups cpu_power.
6807 * cpu_power indicates the capacity of sched group, which is used while
6808 * distributing the load between different sched groups in a sched domain.
6809 * Typically cpu_power for all the groups in a sched domain will be same unless
6810 * there are asymmetries in the topology. If there are asymmetries, group
6811 * having more cpu_power will pickup more load compared to the group having
6814 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6815 * the maximum number of tasks a group can handle in the presence of other idle
6816 * or lightly loaded groups in the same sched domain.
6818 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6820 struct sched_domain *child;
6821 struct sched_group *group;
6823 WARN_ON(!sd || !sd->groups);
6825 if (cpu != first_cpu(sd->groups->cpumask))
6830 sd->groups->__cpu_power = 0;
6833 * For perf policy, if the groups in child domain share resources
6834 * (for example cores sharing some portions of the cache hierarchy
6835 * or SMT), then set this domain groups cpu_power such that each group
6836 * can handle only one task, when there are other idle groups in the
6837 * same sched domain.
6839 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6841 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6842 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6847 * add cpu_power of each child group to this groups cpu_power
6849 group = child->groups;
6851 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6852 group = group->next;
6853 } while (group != child->groups);
6857 * Initializers for schedule domains
6858 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6861 #define SD_INIT(sd, type) sd_init_##type(sd)
6862 #define SD_INIT_FUNC(type) \
6863 static noinline void sd_init_##type(struct sched_domain *sd) \
6865 memset(sd, 0, sizeof(*sd)); \
6866 *sd = SD_##type##_INIT; \
6867 sd->level = SD_LV_##type; \
6872 SD_INIT_FUNC(ALLNODES)
6875 #ifdef CONFIG_SCHED_SMT
6876 SD_INIT_FUNC(SIBLING)
6878 #ifdef CONFIG_SCHED_MC
6883 * To minimize stack usage kmalloc room for cpumasks and share the
6884 * space as the usage in build_sched_domains() dictates. Used only
6885 * if the amount of space is significant.
6888 cpumask_t tmpmask; /* make this one first */
6891 cpumask_t this_sibling_map;
6892 cpumask_t this_core_map;
6894 cpumask_t send_covered;
6897 cpumask_t domainspan;
6899 cpumask_t notcovered;
6904 #define SCHED_CPUMASK_ALLOC 1
6905 #define SCHED_CPUMASK_FREE(v) kfree(v)
6906 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
6908 #define SCHED_CPUMASK_ALLOC 0
6909 #define SCHED_CPUMASK_FREE(v)
6910 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
6913 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
6914 ((unsigned long)(a) + offsetof(struct allmasks, v))
6916 static int default_relax_domain_level = -1;
6918 static int __init setup_relax_domain_level(char *str)
6922 val = simple_strtoul(str, NULL, 0);
6923 if (val < SD_LV_MAX)
6924 default_relax_domain_level = val;
6928 __setup("relax_domain_level=", setup_relax_domain_level);
6930 static void set_domain_attribute(struct sched_domain *sd,
6931 struct sched_domain_attr *attr)
6935 if (!attr || attr->relax_domain_level < 0) {
6936 if (default_relax_domain_level < 0)
6939 request = default_relax_domain_level;
6941 request = attr->relax_domain_level;
6942 if (request < sd->level) {
6943 /* turn off idle balance on this domain */
6944 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
6946 /* turn on idle balance on this domain */
6947 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
6952 * Build sched domains for a given set of cpus and attach the sched domains
6953 * to the individual cpus
6955 static int __build_sched_domains(const cpumask_t *cpu_map,
6956 struct sched_domain_attr *attr)
6959 struct root_domain *rd;
6960 SCHED_CPUMASK_DECLARE(allmasks);
6963 struct sched_group **sched_group_nodes = NULL;
6964 int sd_allnodes = 0;
6967 * Allocate the per-node list of sched groups
6969 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6971 if (!sched_group_nodes) {
6972 printk(KERN_WARNING "Can not alloc sched group node list\n");
6977 rd = alloc_rootdomain();
6979 printk(KERN_WARNING "Cannot alloc root domain\n");
6981 kfree(sched_group_nodes);
6986 #if SCHED_CPUMASK_ALLOC
6987 /* get space for all scratch cpumask variables */
6988 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
6990 printk(KERN_WARNING "Cannot alloc cpumask array\n");
6993 kfree(sched_group_nodes);
6998 tmpmask = (cpumask_t *)allmasks;
7002 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7006 * Set up domains for cpus specified by the cpu_map.
7008 for_each_cpu_mask(i, *cpu_map) {
7009 struct sched_domain *sd = NULL, *p;
7010 SCHED_CPUMASK_VAR(nodemask, allmasks);
7012 *nodemask = node_to_cpumask(cpu_to_node(i));
7013 cpus_and(*nodemask, *nodemask, *cpu_map);
7016 if (cpus_weight(*cpu_map) >
7017 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7018 sd = &per_cpu(allnodes_domains, i);
7019 SD_INIT(sd, ALLNODES);
7020 set_domain_attribute(sd, attr);
7021 sd->span = *cpu_map;
7022 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7028 sd = &per_cpu(node_domains, i);
7030 set_domain_attribute(sd, attr);
7031 sched_domain_node_span(cpu_to_node(i), &sd->span);
7035 cpus_and(sd->span, sd->span, *cpu_map);
7039 sd = &per_cpu(phys_domains, i);
7041 set_domain_attribute(sd, attr);
7042 sd->span = *nodemask;
7046 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7048 #ifdef CONFIG_SCHED_MC
7050 sd = &per_cpu(core_domains, i);
7052 set_domain_attribute(sd, attr);
7053 sd->span = cpu_coregroup_map(i);
7054 cpus_and(sd->span, sd->span, *cpu_map);
7057 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7060 #ifdef CONFIG_SCHED_SMT
7062 sd = &per_cpu(cpu_domains, i);
7063 SD_INIT(sd, SIBLING);
7064 set_domain_attribute(sd, attr);
7065 sd->span = per_cpu(cpu_sibling_map, i);
7066 cpus_and(sd->span, sd->span, *cpu_map);
7069 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7073 #ifdef CONFIG_SCHED_SMT
7074 /* Set up CPU (sibling) groups */
7075 for_each_cpu_mask(i, *cpu_map) {
7076 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7077 SCHED_CPUMASK_VAR(send_covered, allmasks);
7079 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7080 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7081 if (i != first_cpu(*this_sibling_map))
7084 init_sched_build_groups(this_sibling_map, cpu_map,
7086 send_covered, tmpmask);
7090 #ifdef CONFIG_SCHED_MC
7091 /* Set up multi-core groups */
7092 for_each_cpu_mask(i, *cpu_map) {
7093 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7094 SCHED_CPUMASK_VAR(send_covered, allmasks);
7096 *this_core_map = cpu_coregroup_map(i);
7097 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7098 if (i != first_cpu(*this_core_map))
7101 init_sched_build_groups(this_core_map, cpu_map,
7103 send_covered, tmpmask);
7107 /* Set up physical groups */
7108 for (i = 0; i < MAX_NUMNODES; i++) {
7109 SCHED_CPUMASK_VAR(nodemask, allmasks);
7110 SCHED_CPUMASK_VAR(send_covered, allmasks);
7112 *nodemask = node_to_cpumask(i);
7113 cpus_and(*nodemask, *nodemask, *cpu_map);
7114 if (cpus_empty(*nodemask))
7117 init_sched_build_groups(nodemask, cpu_map,
7119 send_covered, tmpmask);
7123 /* Set up node groups */
7125 SCHED_CPUMASK_VAR(send_covered, allmasks);
7127 init_sched_build_groups(cpu_map, cpu_map,
7128 &cpu_to_allnodes_group,
7129 send_covered, tmpmask);
7132 for (i = 0; i < MAX_NUMNODES; i++) {
7133 /* Set up node groups */
7134 struct sched_group *sg, *prev;
7135 SCHED_CPUMASK_VAR(nodemask, allmasks);
7136 SCHED_CPUMASK_VAR(domainspan, allmasks);
7137 SCHED_CPUMASK_VAR(covered, allmasks);
7140 *nodemask = node_to_cpumask(i);
7141 cpus_clear(*covered);
7143 cpus_and(*nodemask, *nodemask, *cpu_map);
7144 if (cpus_empty(*nodemask)) {
7145 sched_group_nodes[i] = NULL;
7149 sched_domain_node_span(i, domainspan);
7150 cpus_and(*domainspan, *domainspan, *cpu_map);
7152 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7154 printk(KERN_WARNING "Can not alloc domain group for "
7158 sched_group_nodes[i] = sg;
7159 for_each_cpu_mask(j, *nodemask) {
7160 struct sched_domain *sd;
7162 sd = &per_cpu(node_domains, j);
7165 sg->__cpu_power = 0;
7166 sg->cpumask = *nodemask;
7168 cpus_or(*covered, *covered, *nodemask);
7171 for (j = 0; j < MAX_NUMNODES; j++) {
7172 SCHED_CPUMASK_VAR(notcovered, allmasks);
7173 int n = (i + j) % MAX_NUMNODES;
7174 node_to_cpumask_ptr(pnodemask, n);
7176 cpus_complement(*notcovered, *covered);
7177 cpus_and(*tmpmask, *notcovered, *cpu_map);
7178 cpus_and(*tmpmask, *tmpmask, *domainspan);
7179 if (cpus_empty(*tmpmask))
7182 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7183 if (cpus_empty(*tmpmask))
7186 sg = kmalloc_node(sizeof(struct sched_group),
7190 "Can not alloc domain group for node %d\n", j);
7193 sg->__cpu_power = 0;
7194 sg->cpumask = *tmpmask;
7195 sg->next = prev->next;
7196 cpus_or(*covered, *covered, *tmpmask);
7203 /* Calculate CPU power for physical packages and nodes */
7204 #ifdef CONFIG_SCHED_SMT
7205 for_each_cpu_mask(i, *cpu_map) {
7206 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7208 init_sched_groups_power(i, sd);
7211 #ifdef CONFIG_SCHED_MC
7212 for_each_cpu_mask(i, *cpu_map) {
7213 struct sched_domain *sd = &per_cpu(core_domains, i);
7215 init_sched_groups_power(i, sd);
7219 for_each_cpu_mask(i, *cpu_map) {
7220 struct sched_domain *sd = &per_cpu(phys_domains, i);
7222 init_sched_groups_power(i, sd);
7226 for (i = 0; i < MAX_NUMNODES; i++)
7227 init_numa_sched_groups_power(sched_group_nodes[i]);
7230 struct sched_group *sg;
7232 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7234 init_numa_sched_groups_power(sg);
7238 /* Attach the domains */
7239 for_each_cpu_mask(i, *cpu_map) {
7240 struct sched_domain *sd;
7241 #ifdef CONFIG_SCHED_SMT
7242 sd = &per_cpu(cpu_domains, i);
7243 #elif defined(CONFIG_SCHED_MC)
7244 sd = &per_cpu(core_domains, i);
7246 sd = &per_cpu(phys_domains, i);
7248 cpu_attach_domain(sd, rd, i);
7251 SCHED_CPUMASK_FREE((void *)allmasks);
7256 free_sched_groups(cpu_map, tmpmask);
7257 SCHED_CPUMASK_FREE((void *)allmasks);
7262 static int build_sched_domains(const cpumask_t *cpu_map)
7264 return __build_sched_domains(cpu_map, NULL);
7267 static cpumask_t *doms_cur; /* current sched domains */
7268 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7269 static struct sched_domain_attr *dattr_cur;
7270 /* attribues of custom domains in 'doms_cur' */
7273 * Special case: If a kmalloc of a doms_cur partition (array of
7274 * cpumask_t) fails, then fallback to a single sched domain,
7275 * as determined by the single cpumask_t fallback_doms.
7277 static cpumask_t fallback_doms;
7279 void __attribute__((weak)) arch_update_cpu_topology(void)
7284 * Free current domain masks.
7285 * Called after all cpus are attached to NULL domain.
7287 static void free_sched_domains(void)
7290 if (doms_cur != &fallback_doms)
7292 doms_cur = &fallback_doms;
7296 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7297 * For now this just excludes isolated cpus, but could be used to
7298 * exclude other special cases in the future.
7300 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7304 arch_update_cpu_topology();
7306 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7308 doms_cur = &fallback_doms;
7309 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7311 err = build_sched_domains(doms_cur);
7312 register_sched_domain_sysctl();
7317 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7320 free_sched_groups(cpu_map, tmpmask);
7324 * Detach sched domains from a group of cpus specified in cpu_map
7325 * These cpus will now be attached to the NULL domain
7327 static void detach_destroy_domains(const cpumask_t *cpu_map)
7332 unregister_sched_domain_sysctl();
7334 for_each_cpu_mask(i, *cpu_map)
7335 cpu_attach_domain(NULL, &def_root_domain, i);
7336 synchronize_sched();
7337 arch_destroy_sched_domains(cpu_map, &tmpmask);
7340 /* handle null as "default" */
7341 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7342 struct sched_domain_attr *new, int idx_new)
7344 struct sched_domain_attr tmp;
7351 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7352 new ? (new + idx_new) : &tmp,
7353 sizeof(struct sched_domain_attr));
7357 * Partition sched domains as specified by the 'ndoms_new'
7358 * cpumasks in the array doms_new[] of cpumasks. This compares
7359 * doms_new[] to the current sched domain partitioning, doms_cur[].
7360 * It destroys each deleted domain and builds each new domain.
7362 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7363 * The masks don't intersect (don't overlap.) We should setup one
7364 * sched domain for each mask. CPUs not in any of the cpumasks will
7365 * not be load balanced. If the same cpumask appears both in the
7366 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7369 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7370 * ownership of it and will kfree it when done with it. If the caller
7371 * failed the kmalloc call, then it can pass in doms_new == NULL,
7372 * and partition_sched_domains() will fallback to the single partition
7375 * Call with hotplug lock held
7377 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7378 struct sched_domain_attr *dattr_new)
7382 mutex_lock(&sched_domains_mutex);
7384 /* always unregister in case we don't destroy any domains */
7385 unregister_sched_domain_sysctl();
7387 if (doms_new == NULL) {
7389 doms_new = &fallback_doms;
7390 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7394 /* Destroy deleted domains */
7395 for (i = 0; i < ndoms_cur; i++) {
7396 for (j = 0; j < ndoms_new; j++) {
7397 if (cpus_equal(doms_cur[i], doms_new[j])
7398 && dattrs_equal(dattr_cur, i, dattr_new, j))
7401 /* no match - a current sched domain not in new doms_new[] */
7402 detach_destroy_domains(doms_cur + i);
7407 /* Build new domains */
7408 for (i = 0; i < ndoms_new; i++) {
7409 for (j = 0; j < ndoms_cur; j++) {
7410 if (cpus_equal(doms_new[i], doms_cur[j])
7411 && dattrs_equal(dattr_new, i, dattr_cur, j))
7414 /* no match - add a new doms_new */
7415 __build_sched_domains(doms_new + i,
7416 dattr_new ? dattr_new + i : NULL);
7421 /* Remember the new sched domains */
7422 if (doms_cur != &fallback_doms)
7424 kfree(dattr_cur); /* kfree(NULL) is safe */
7425 doms_cur = doms_new;
7426 dattr_cur = dattr_new;
7427 ndoms_cur = ndoms_new;
7429 register_sched_domain_sysctl();
7431 mutex_unlock(&sched_domains_mutex);
7434 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7435 int arch_reinit_sched_domains(void)
7440 mutex_lock(&sched_domains_mutex);
7441 detach_destroy_domains(&cpu_online_map);
7442 free_sched_domains();
7443 err = arch_init_sched_domains(&cpu_online_map);
7444 mutex_unlock(&sched_domains_mutex);
7450 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7454 if (buf[0] != '0' && buf[0] != '1')
7458 sched_smt_power_savings = (buf[0] == '1');
7460 sched_mc_power_savings = (buf[0] == '1');
7462 ret = arch_reinit_sched_domains();
7464 return ret ? ret : count;
7467 #ifdef CONFIG_SCHED_MC
7468 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7470 return sprintf(page, "%u\n", sched_mc_power_savings);
7472 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7473 const char *buf, size_t count)
7475 return sched_power_savings_store(buf, count, 0);
7477 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7478 sched_mc_power_savings_store);
7481 #ifdef CONFIG_SCHED_SMT
7482 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7484 return sprintf(page, "%u\n", sched_smt_power_savings);
7486 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7487 const char *buf, size_t count)
7489 return sched_power_savings_store(buf, count, 1);
7491 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7492 sched_smt_power_savings_store);
7495 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7499 #ifdef CONFIG_SCHED_SMT
7501 err = sysfs_create_file(&cls->kset.kobj,
7502 &attr_sched_smt_power_savings.attr);
7504 #ifdef CONFIG_SCHED_MC
7505 if (!err && mc_capable())
7506 err = sysfs_create_file(&cls->kset.kobj,
7507 &attr_sched_mc_power_savings.attr);
7511 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7514 * Force a reinitialization of the sched domains hierarchy. The domains
7515 * and groups cannot be updated in place without racing with the balancing
7516 * code, so we temporarily attach all running cpus to the NULL domain
7517 * which will prevent rebalancing while the sched domains are recalculated.
7519 static int update_sched_domains(struct notifier_block *nfb,
7520 unsigned long action, void *hcpu)
7522 int cpu = (int)(long)hcpu;
7525 case CPU_DOWN_PREPARE:
7526 case CPU_DOWN_PREPARE_FROZEN:
7527 disable_runtime(cpu_rq(cpu));
7529 case CPU_UP_PREPARE:
7530 case CPU_UP_PREPARE_FROZEN:
7531 detach_destroy_domains(&cpu_online_map);
7532 free_sched_domains();
7536 case CPU_DOWN_FAILED:
7537 case CPU_DOWN_FAILED_FROZEN:
7539 case CPU_ONLINE_FROZEN:
7540 enable_runtime(cpu_rq(cpu));
7542 case CPU_UP_CANCELED:
7543 case CPU_UP_CANCELED_FROZEN:
7545 case CPU_DEAD_FROZEN:
7547 * Fall through and re-initialise the domains.
7554 #ifndef CONFIG_CPUSETS
7556 * Create default domain partitioning if cpusets are disabled.
7557 * Otherwise we let cpusets rebuild the domains based on the
7561 /* The hotplug lock is already held by cpu_up/cpu_down */
7562 arch_init_sched_domains(&cpu_online_map);
7568 void __init sched_init_smp(void)
7570 cpumask_t non_isolated_cpus;
7572 #if defined(CONFIG_NUMA)
7573 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7575 BUG_ON(sched_group_nodes_bycpu == NULL);
7578 mutex_lock(&sched_domains_mutex);
7579 arch_init_sched_domains(&cpu_online_map);
7580 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7581 if (cpus_empty(non_isolated_cpus))
7582 cpu_set(smp_processor_id(), non_isolated_cpus);
7583 mutex_unlock(&sched_domains_mutex);
7585 /* XXX: Theoretical race here - CPU may be hotplugged now */
7586 hotcpu_notifier(update_sched_domains, 0);
7589 /* Move init over to a non-isolated CPU */
7590 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7592 sched_init_granularity();
7595 void __init sched_init_smp(void)
7597 sched_init_granularity();
7599 #endif /* CONFIG_SMP */
7601 int in_sched_functions(unsigned long addr)
7603 return in_lock_functions(addr) ||
7604 (addr >= (unsigned long)__sched_text_start
7605 && addr < (unsigned long)__sched_text_end);
7608 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7610 cfs_rq->tasks_timeline = RB_ROOT;
7611 INIT_LIST_HEAD(&cfs_rq->tasks);
7612 #ifdef CONFIG_FAIR_GROUP_SCHED
7615 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7618 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7620 struct rt_prio_array *array;
7623 array = &rt_rq->active;
7624 for (i = 0; i < MAX_RT_PRIO; i++) {
7625 INIT_LIST_HEAD(array->queue + i);
7626 __clear_bit(i, array->bitmap);
7628 /* delimiter for bitsearch: */
7629 __set_bit(MAX_RT_PRIO, array->bitmap);
7631 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7632 rt_rq->highest_prio = MAX_RT_PRIO;
7635 rt_rq->rt_nr_migratory = 0;
7636 rt_rq->overloaded = 0;
7640 rt_rq->rt_throttled = 0;
7641 rt_rq->rt_runtime = 0;
7642 spin_lock_init(&rt_rq->rt_runtime_lock);
7644 #ifdef CONFIG_RT_GROUP_SCHED
7645 rt_rq->rt_nr_boosted = 0;
7650 #ifdef CONFIG_FAIR_GROUP_SCHED
7651 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7652 struct sched_entity *se, int cpu, int add,
7653 struct sched_entity *parent)
7655 struct rq *rq = cpu_rq(cpu);
7656 tg->cfs_rq[cpu] = cfs_rq;
7657 init_cfs_rq(cfs_rq, rq);
7660 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7663 /* se could be NULL for init_task_group */
7668 se->cfs_rq = &rq->cfs;
7670 se->cfs_rq = parent->my_q;
7673 se->load.weight = tg->shares;
7674 se->load.inv_weight = 0;
7675 se->parent = parent;
7679 #ifdef CONFIG_RT_GROUP_SCHED
7680 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7681 struct sched_rt_entity *rt_se, int cpu, int add,
7682 struct sched_rt_entity *parent)
7684 struct rq *rq = cpu_rq(cpu);
7686 tg->rt_rq[cpu] = rt_rq;
7687 init_rt_rq(rt_rq, rq);
7689 rt_rq->rt_se = rt_se;
7690 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7692 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7694 tg->rt_se[cpu] = rt_se;
7699 rt_se->rt_rq = &rq->rt;
7701 rt_se->rt_rq = parent->my_q;
7703 rt_se->my_q = rt_rq;
7704 rt_se->parent = parent;
7705 INIT_LIST_HEAD(&rt_se->run_list);
7709 void __init sched_init(void)
7712 unsigned long alloc_size = 0, ptr;
7714 #ifdef CONFIG_FAIR_GROUP_SCHED
7715 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7717 #ifdef CONFIG_RT_GROUP_SCHED
7718 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7720 #ifdef CONFIG_USER_SCHED
7724 * As sched_init() is called before page_alloc is setup,
7725 * we use alloc_bootmem().
7728 ptr = (unsigned long)alloc_bootmem(alloc_size);
7730 #ifdef CONFIG_FAIR_GROUP_SCHED
7731 init_task_group.se = (struct sched_entity **)ptr;
7732 ptr += nr_cpu_ids * sizeof(void **);
7734 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7735 ptr += nr_cpu_ids * sizeof(void **);
7737 #ifdef CONFIG_USER_SCHED
7738 root_task_group.se = (struct sched_entity **)ptr;
7739 ptr += nr_cpu_ids * sizeof(void **);
7741 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7742 ptr += nr_cpu_ids * sizeof(void **);
7743 #endif /* CONFIG_USER_SCHED */
7744 #endif /* CONFIG_FAIR_GROUP_SCHED */
7745 #ifdef CONFIG_RT_GROUP_SCHED
7746 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7747 ptr += nr_cpu_ids * sizeof(void **);
7749 init_task_group.rt_rq = (struct rt_rq **)ptr;
7750 ptr += nr_cpu_ids * sizeof(void **);
7752 #ifdef CONFIG_USER_SCHED
7753 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7754 ptr += nr_cpu_ids * sizeof(void **);
7756 root_task_group.rt_rq = (struct rt_rq **)ptr;
7757 ptr += nr_cpu_ids * sizeof(void **);
7758 #endif /* CONFIG_USER_SCHED */
7759 #endif /* CONFIG_RT_GROUP_SCHED */
7763 init_defrootdomain();
7766 init_rt_bandwidth(&def_rt_bandwidth,
7767 global_rt_period(), global_rt_runtime());
7769 #ifdef CONFIG_RT_GROUP_SCHED
7770 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7771 global_rt_period(), global_rt_runtime());
7772 #ifdef CONFIG_USER_SCHED
7773 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7774 global_rt_period(), RUNTIME_INF);
7775 #endif /* CONFIG_USER_SCHED */
7776 #endif /* CONFIG_RT_GROUP_SCHED */
7778 #ifdef CONFIG_GROUP_SCHED
7779 list_add(&init_task_group.list, &task_groups);
7780 INIT_LIST_HEAD(&init_task_group.children);
7782 #ifdef CONFIG_USER_SCHED
7783 INIT_LIST_HEAD(&root_task_group.children);
7784 init_task_group.parent = &root_task_group;
7785 list_add(&init_task_group.siblings, &root_task_group.children);
7786 #endif /* CONFIG_USER_SCHED */
7787 #endif /* CONFIG_GROUP_SCHED */
7789 for_each_possible_cpu(i) {
7793 spin_lock_init(&rq->lock);
7794 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7796 init_cfs_rq(&rq->cfs, rq);
7797 init_rt_rq(&rq->rt, rq);
7798 #ifdef CONFIG_FAIR_GROUP_SCHED
7799 init_task_group.shares = init_task_group_load;
7800 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7801 #ifdef CONFIG_CGROUP_SCHED
7803 * How much cpu bandwidth does init_task_group get?
7805 * In case of task-groups formed thr' the cgroup filesystem, it
7806 * gets 100% of the cpu resources in the system. This overall
7807 * system cpu resource is divided among the tasks of
7808 * init_task_group and its child task-groups in a fair manner,
7809 * based on each entity's (task or task-group's) weight
7810 * (se->load.weight).
7812 * In other words, if init_task_group has 10 tasks of weight
7813 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7814 * then A0's share of the cpu resource is:
7816 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7818 * We achieve this by letting init_task_group's tasks sit
7819 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7821 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7822 #elif defined CONFIG_USER_SCHED
7823 root_task_group.shares = NICE_0_LOAD;
7824 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
7826 * In case of task-groups formed thr' the user id of tasks,
7827 * init_task_group represents tasks belonging to root user.
7828 * Hence it forms a sibling of all subsequent groups formed.
7829 * In this case, init_task_group gets only a fraction of overall
7830 * system cpu resource, based on the weight assigned to root
7831 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7832 * by letting tasks of init_task_group sit in a separate cfs_rq
7833 * (init_cfs_rq) and having one entity represent this group of
7834 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7836 init_tg_cfs_entry(&init_task_group,
7837 &per_cpu(init_cfs_rq, i),
7838 &per_cpu(init_sched_entity, i), i, 1,
7839 root_task_group.se[i]);
7842 #endif /* CONFIG_FAIR_GROUP_SCHED */
7844 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7845 #ifdef CONFIG_RT_GROUP_SCHED
7846 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7847 #ifdef CONFIG_CGROUP_SCHED
7848 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7849 #elif defined CONFIG_USER_SCHED
7850 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
7851 init_tg_rt_entry(&init_task_group,
7852 &per_cpu(init_rt_rq, i),
7853 &per_cpu(init_sched_rt_entity, i), i, 1,
7854 root_task_group.rt_se[i]);
7858 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7859 rq->cpu_load[j] = 0;
7863 rq->active_balance = 0;
7864 rq->next_balance = jiffies;
7868 rq->migration_thread = NULL;
7869 INIT_LIST_HEAD(&rq->migration_queue);
7870 rq_attach_root(rq, &def_root_domain);
7873 atomic_set(&rq->nr_iowait, 0);
7876 set_load_weight(&init_task);
7878 #ifdef CONFIG_PREEMPT_NOTIFIERS
7879 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7883 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7886 #ifdef CONFIG_RT_MUTEXES
7887 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7891 * The boot idle thread does lazy MMU switching as well:
7893 atomic_inc(&init_mm.mm_count);
7894 enter_lazy_tlb(&init_mm, current);
7897 * Make us the idle thread. Technically, schedule() should not be
7898 * called from this thread, however somewhere below it might be,
7899 * but because we are the idle thread, we just pick up running again
7900 * when this runqueue becomes "idle".
7902 init_idle(current, smp_processor_id());
7904 * During early bootup we pretend to be a normal task:
7906 current->sched_class = &fair_sched_class;
7908 scheduler_running = 1;
7911 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7912 void __might_sleep(char *file, int line)
7915 static unsigned long prev_jiffy; /* ratelimiting */
7917 if ((in_atomic() || irqs_disabled()) &&
7918 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7919 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7921 prev_jiffy = jiffies;
7922 printk(KERN_ERR "BUG: sleeping function called from invalid"
7923 " context at %s:%d\n", file, line);
7924 printk("in_atomic():%d, irqs_disabled():%d\n",
7925 in_atomic(), irqs_disabled());
7926 debug_show_held_locks(current);
7927 if (irqs_disabled())
7928 print_irqtrace_events(current);
7933 EXPORT_SYMBOL(__might_sleep);
7936 #ifdef CONFIG_MAGIC_SYSRQ
7937 static void normalize_task(struct rq *rq, struct task_struct *p)
7941 update_rq_clock(rq);
7942 on_rq = p->se.on_rq;
7944 deactivate_task(rq, p, 0);
7945 __setscheduler(rq, p, SCHED_NORMAL, 0);
7947 activate_task(rq, p, 0);
7948 resched_task(rq->curr);
7952 void normalize_rt_tasks(void)
7954 struct task_struct *g, *p;
7955 unsigned long flags;
7958 read_lock_irqsave(&tasklist_lock, flags);
7959 do_each_thread(g, p) {
7961 * Only normalize user tasks:
7966 p->se.exec_start = 0;
7967 #ifdef CONFIG_SCHEDSTATS
7968 p->se.wait_start = 0;
7969 p->se.sleep_start = 0;
7970 p->se.block_start = 0;
7975 * Renice negative nice level userspace
7978 if (TASK_NICE(p) < 0 && p->mm)
7979 set_user_nice(p, 0);
7983 spin_lock(&p->pi_lock);
7984 rq = __task_rq_lock(p);
7986 normalize_task(rq, p);
7988 __task_rq_unlock(rq);
7989 spin_unlock(&p->pi_lock);
7990 } while_each_thread(g, p);
7992 read_unlock_irqrestore(&tasklist_lock, flags);
7995 #endif /* CONFIG_MAGIC_SYSRQ */
7999 * These functions are only useful for the IA64 MCA handling.
8001 * They can only be called when the whole system has been
8002 * stopped - every CPU needs to be quiescent, and no scheduling
8003 * activity can take place. Using them for anything else would
8004 * be a serious bug, and as a result, they aren't even visible
8005 * under any other configuration.
8009 * curr_task - return the current task for a given cpu.
8010 * @cpu: the processor in question.
8012 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8014 struct task_struct *curr_task(int cpu)
8016 return cpu_curr(cpu);
8020 * set_curr_task - set the current task for a given cpu.
8021 * @cpu: the processor in question.
8022 * @p: the task pointer to set.
8024 * Description: This function must only be used when non-maskable interrupts
8025 * are serviced on a separate stack. It allows the architecture to switch the
8026 * notion of the current task on a cpu in a non-blocking manner. This function
8027 * must be called with all CPU's synchronized, and interrupts disabled, the
8028 * and caller must save the original value of the current task (see
8029 * curr_task() above) and restore that value before reenabling interrupts and
8030 * re-starting the system.
8032 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8034 void set_curr_task(int cpu, struct task_struct *p)
8041 #ifdef CONFIG_FAIR_GROUP_SCHED
8042 static void free_fair_sched_group(struct task_group *tg)
8046 for_each_possible_cpu(i) {
8048 kfree(tg->cfs_rq[i]);
8058 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8060 struct cfs_rq *cfs_rq;
8061 struct sched_entity *se, *parent_se;
8065 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8068 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8072 tg->shares = NICE_0_LOAD;
8074 for_each_possible_cpu(i) {
8077 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8078 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8082 se = kmalloc_node(sizeof(struct sched_entity),
8083 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8087 parent_se = parent ? parent->se[i] : NULL;
8088 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8097 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8099 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8100 &cpu_rq(cpu)->leaf_cfs_rq_list);
8103 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8105 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8107 #else /* !CONFG_FAIR_GROUP_SCHED */
8108 static inline void free_fair_sched_group(struct task_group *tg)
8113 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8118 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8122 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8125 #endif /* CONFIG_FAIR_GROUP_SCHED */
8127 #ifdef CONFIG_RT_GROUP_SCHED
8128 static void free_rt_sched_group(struct task_group *tg)
8132 destroy_rt_bandwidth(&tg->rt_bandwidth);
8134 for_each_possible_cpu(i) {
8136 kfree(tg->rt_rq[i]);
8138 kfree(tg->rt_se[i]);
8146 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8148 struct rt_rq *rt_rq;
8149 struct sched_rt_entity *rt_se, *parent_se;
8153 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8156 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8160 init_rt_bandwidth(&tg->rt_bandwidth,
8161 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8163 for_each_possible_cpu(i) {
8166 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8167 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8171 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8172 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8176 parent_se = parent ? parent->rt_se[i] : NULL;
8177 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8186 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8188 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8189 &cpu_rq(cpu)->leaf_rt_rq_list);
8192 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8194 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8196 #else /* !CONFIG_RT_GROUP_SCHED */
8197 static inline void free_rt_sched_group(struct task_group *tg)
8202 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8207 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8211 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8214 #endif /* CONFIG_RT_GROUP_SCHED */
8216 #ifdef CONFIG_GROUP_SCHED
8217 static void free_sched_group(struct task_group *tg)
8219 free_fair_sched_group(tg);
8220 free_rt_sched_group(tg);
8224 /* allocate runqueue etc for a new task group */
8225 struct task_group *sched_create_group(struct task_group *parent)
8227 struct task_group *tg;
8228 unsigned long flags;
8231 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8233 return ERR_PTR(-ENOMEM);
8235 if (!alloc_fair_sched_group(tg, parent))
8238 if (!alloc_rt_sched_group(tg, parent))
8241 spin_lock_irqsave(&task_group_lock, flags);
8242 for_each_possible_cpu(i) {
8243 register_fair_sched_group(tg, i);
8244 register_rt_sched_group(tg, i);
8246 list_add_rcu(&tg->list, &task_groups);
8248 WARN_ON(!parent); /* root should already exist */
8250 tg->parent = parent;
8251 list_add_rcu(&tg->siblings, &parent->children);
8252 INIT_LIST_HEAD(&tg->children);
8253 spin_unlock_irqrestore(&task_group_lock, flags);
8258 free_sched_group(tg);
8259 return ERR_PTR(-ENOMEM);
8262 /* rcu callback to free various structures associated with a task group */
8263 static void free_sched_group_rcu(struct rcu_head *rhp)
8265 /* now it should be safe to free those cfs_rqs */
8266 free_sched_group(container_of(rhp, struct task_group, rcu));
8269 /* Destroy runqueue etc associated with a task group */
8270 void sched_destroy_group(struct task_group *tg)
8272 unsigned long flags;
8275 spin_lock_irqsave(&task_group_lock, flags);
8276 for_each_possible_cpu(i) {
8277 unregister_fair_sched_group(tg, i);
8278 unregister_rt_sched_group(tg, i);
8280 list_del_rcu(&tg->list);
8281 list_del_rcu(&tg->siblings);
8282 spin_unlock_irqrestore(&task_group_lock, flags);
8284 /* wait for possible concurrent references to cfs_rqs complete */
8285 call_rcu(&tg->rcu, free_sched_group_rcu);
8288 /* change task's runqueue when it moves between groups.
8289 * The caller of this function should have put the task in its new group
8290 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8291 * reflect its new group.
8293 void sched_move_task(struct task_struct *tsk)
8296 unsigned long flags;
8299 rq = task_rq_lock(tsk, &flags);
8301 update_rq_clock(rq);
8303 running = task_current(rq, tsk);
8304 on_rq = tsk->se.on_rq;
8307 dequeue_task(rq, tsk, 0);
8308 if (unlikely(running))
8309 tsk->sched_class->put_prev_task(rq, tsk);
8311 set_task_rq(tsk, task_cpu(tsk));
8313 #ifdef CONFIG_FAIR_GROUP_SCHED
8314 if (tsk->sched_class->moved_group)
8315 tsk->sched_class->moved_group(tsk);
8318 if (unlikely(running))
8319 tsk->sched_class->set_curr_task(rq);
8321 enqueue_task(rq, tsk, 0);
8323 task_rq_unlock(rq, &flags);
8325 #endif /* CONFIG_GROUP_SCHED */
8327 #ifdef CONFIG_FAIR_GROUP_SCHED
8328 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8330 struct cfs_rq *cfs_rq = se->cfs_rq;
8331 struct rq *rq = cfs_rq->rq;
8334 spin_lock_irq(&rq->lock);
8338 dequeue_entity(cfs_rq, se, 0);
8340 se->load.weight = shares;
8341 se->load.inv_weight = 0;
8344 enqueue_entity(cfs_rq, se, 0);
8346 spin_unlock_irq(&rq->lock);
8349 static DEFINE_MUTEX(shares_mutex);
8351 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8354 unsigned long flags;
8357 * We can't change the weight of the root cgroup.
8362 if (shares < MIN_SHARES)
8363 shares = MIN_SHARES;
8364 else if (shares > MAX_SHARES)
8365 shares = MAX_SHARES;
8367 mutex_lock(&shares_mutex);
8368 if (tg->shares == shares)
8371 spin_lock_irqsave(&task_group_lock, flags);
8372 for_each_possible_cpu(i)
8373 unregister_fair_sched_group(tg, i);
8374 list_del_rcu(&tg->siblings);
8375 spin_unlock_irqrestore(&task_group_lock, flags);
8377 /* wait for any ongoing reference to this group to finish */
8378 synchronize_sched();
8381 * Now we are free to modify the group's share on each cpu
8382 * w/o tripping rebalance_share or load_balance_fair.
8384 tg->shares = shares;
8385 for_each_possible_cpu(i)
8386 set_se_shares(tg->se[i], shares);
8389 * Enable load balance activity on this group, by inserting it back on
8390 * each cpu's rq->leaf_cfs_rq_list.
8392 spin_lock_irqsave(&task_group_lock, flags);
8393 for_each_possible_cpu(i)
8394 register_fair_sched_group(tg, i);
8395 list_add_rcu(&tg->siblings, &tg->parent->children);
8396 spin_unlock_irqrestore(&task_group_lock, flags);
8398 mutex_unlock(&shares_mutex);
8402 unsigned long sched_group_shares(struct task_group *tg)
8408 #ifdef CONFIG_RT_GROUP_SCHED
8410 * Ensure that the real time constraints are schedulable.
8412 static DEFINE_MUTEX(rt_constraints_mutex);
8414 static unsigned long to_ratio(u64 period, u64 runtime)
8416 if (runtime == RUNTIME_INF)
8419 return div64_u64(runtime << 16, period);
8422 #ifdef CONFIG_CGROUP_SCHED
8423 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8425 struct task_group *tgi, *parent = tg->parent;
8426 unsigned long total = 0;
8429 if (global_rt_period() < period)
8432 return to_ratio(period, runtime) <
8433 to_ratio(global_rt_period(), global_rt_runtime());
8436 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8440 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8444 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8445 tgi->rt_bandwidth.rt_runtime);
8449 return total + to_ratio(period, runtime) <=
8450 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8451 parent->rt_bandwidth.rt_runtime);
8453 #elif defined CONFIG_USER_SCHED
8454 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8456 struct task_group *tgi;
8457 unsigned long total = 0;
8458 unsigned long global_ratio =
8459 to_ratio(global_rt_period(), global_rt_runtime());
8462 list_for_each_entry_rcu(tgi, &task_groups, list) {
8466 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8467 tgi->rt_bandwidth.rt_runtime);
8471 return total + to_ratio(period, runtime) < global_ratio;
8475 /* Must be called with tasklist_lock held */
8476 static inline int tg_has_rt_tasks(struct task_group *tg)
8478 struct task_struct *g, *p;
8479 do_each_thread(g, p) {
8480 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8482 } while_each_thread(g, p);
8486 static int tg_set_bandwidth(struct task_group *tg,
8487 u64 rt_period, u64 rt_runtime)
8491 mutex_lock(&rt_constraints_mutex);
8492 read_lock(&tasklist_lock);
8493 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8497 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8502 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8503 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8504 tg->rt_bandwidth.rt_runtime = rt_runtime;
8506 for_each_possible_cpu(i) {
8507 struct rt_rq *rt_rq = tg->rt_rq[i];
8509 spin_lock(&rt_rq->rt_runtime_lock);
8510 rt_rq->rt_runtime = rt_runtime;
8511 spin_unlock(&rt_rq->rt_runtime_lock);
8513 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8515 read_unlock(&tasklist_lock);
8516 mutex_unlock(&rt_constraints_mutex);
8521 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8523 u64 rt_runtime, rt_period;
8525 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8526 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8527 if (rt_runtime_us < 0)
8528 rt_runtime = RUNTIME_INF;
8530 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8533 long sched_group_rt_runtime(struct task_group *tg)
8537 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8540 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8541 do_div(rt_runtime_us, NSEC_PER_USEC);
8542 return rt_runtime_us;
8545 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8547 u64 rt_runtime, rt_period;
8549 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8550 rt_runtime = tg->rt_bandwidth.rt_runtime;
8552 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8555 long sched_group_rt_period(struct task_group *tg)
8559 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8560 do_div(rt_period_us, NSEC_PER_USEC);
8561 return rt_period_us;
8564 static int sched_rt_global_constraints(void)
8566 struct task_group *tg = &root_task_group;
8567 u64 rt_runtime, rt_period;
8570 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8571 rt_runtime = tg->rt_bandwidth.rt_runtime;
8573 mutex_lock(&rt_constraints_mutex);
8574 if (!__rt_schedulable(tg, rt_period, rt_runtime))
8576 mutex_unlock(&rt_constraints_mutex);
8580 #else /* !CONFIG_RT_GROUP_SCHED */
8581 static int sched_rt_global_constraints(void)
8583 unsigned long flags;
8586 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8587 for_each_possible_cpu(i) {
8588 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8590 spin_lock(&rt_rq->rt_runtime_lock);
8591 rt_rq->rt_runtime = global_rt_runtime();
8592 spin_unlock(&rt_rq->rt_runtime_lock);
8594 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8598 #endif /* CONFIG_RT_GROUP_SCHED */
8600 int sched_rt_handler(struct ctl_table *table, int write,
8601 struct file *filp, void __user *buffer, size_t *lenp,
8605 int old_period, old_runtime;
8606 static DEFINE_MUTEX(mutex);
8609 old_period = sysctl_sched_rt_period;
8610 old_runtime = sysctl_sched_rt_runtime;
8612 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8614 if (!ret && write) {
8615 ret = sched_rt_global_constraints();
8617 sysctl_sched_rt_period = old_period;
8618 sysctl_sched_rt_runtime = old_runtime;
8620 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8621 def_rt_bandwidth.rt_period =
8622 ns_to_ktime(global_rt_period());
8625 mutex_unlock(&mutex);
8630 #ifdef CONFIG_CGROUP_SCHED
8632 /* return corresponding task_group object of a cgroup */
8633 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8635 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8636 struct task_group, css);
8639 static struct cgroup_subsys_state *
8640 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8642 struct task_group *tg, *parent;
8644 if (!cgrp->parent) {
8645 /* This is early initialization for the top cgroup */
8646 init_task_group.css.cgroup = cgrp;
8647 return &init_task_group.css;
8650 parent = cgroup_tg(cgrp->parent);
8651 tg = sched_create_group(parent);
8653 return ERR_PTR(-ENOMEM);
8655 /* Bind the cgroup to task_group object we just created */
8656 tg->css.cgroup = cgrp;
8662 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8664 struct task_group *tg = cgroup_tg(cgrp);
8666 sched_destroy_group(tg);
8670 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8671 struct task_struct *tsk)
8673 #ifdef CONFIG_RT_GROUP_SCHED
8674 /* Don't accept realtime tasks when there is no way for them to run */
8675 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8678 /* We don't support RT-tasks being in separate groups */
8679 if (tsk->sched_class != &fair_sched_class)
8687 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8688 struct cgroup *old_cont, struct task_struct *tsk)
8690 sched_move_task(tsk);
8693 #ifdef CONFIG_FAIR_GROUP_SCHED
8694 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8697 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8700 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8702 struct task_group *tg = cgroup_tg(cgrp);
8704 return (u64) tg->shares;
8706 #endif /* CONFIG_FAIR_GROUP_SCHED */
8708 #ifdef CONFIG_RT_GROUP_SCHED
8709 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8712 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8715 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8717 return sched_group_rt_runtime(cgroup_tg(cgrp));
8720 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8723 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8726 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8728 return sched_group_rt_period(cgroup_tg(cgrp));
8730 #endif /* CONFIG_RT_GROUP_SCHED */
8732 static struct cftype cpu_files[] = {
8733 #ifdef CONFIG_FAIR_GROUP_SCHED
8736 .read_u64 = cpu_shares_read_u64,
8737 .write_u64 = cpu_shares_write_u64,
8740 #ifdef CONFIG_RT_GROUP_SCHED
8742 .name = "rt_runtime_us",
8743 .read_s64 = cpu_rt_runtime_read,
8744 .write_s64 = cpu_rt_runtime_write,
8747 .name = "rt_period_us",
8748 .read_u64 = cpu_rt_period_read_uint,
8749 .write_u64 = cpu_rt_period_write_uint,
8754 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8756 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8759 struct cgroup_subsys cpu_cgroup_subsys = {
8761 .create = cpu_cgroup_create,
8762 .destroy = cpu_cgroup_destroy,
8763 .can_attach = cpu_cgroup_can_attach,
8764 .attach = cpu_cgroup_attach,
8765 .populate = cpu_cgroup_populate,
8766 .subsys_id = cpu_cgroup_subsys_id,
8770 #endif /* CONFIG_CGROUP_SCHED */
8772 #ifdef CONFIG_CGROUP_CPUACCT
8775 * CPU accounting code for task groups.
8777 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8778 * (balbir@in.ibm.com).
8781 /* track cpu usage of a group of tasks */
8783 struct cgroup_subsys_state css;
8784 /* cpuusage holds pointer to a u64-type object on every cpu */
8788 struct cgroup_subsys cpuacct_subsys;
8790 /* return cpu accounting group corresponding to this container */
8791 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8793 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8794 struct cpuacct, css);
8797 /* return cpu accounting group to which this task belongs */
8798 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8800 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8801 struct cpuacct, css);
8804 /* create a new cpu accounting group */
8805 static struct cgroup_subsys_state *cpuacct_create(
8806 struct cgroup_subsys *ss, struct cgroup *cgrp)
8808 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8811 return ERR_PTR(-ENOMEM);
8813 ca->cpuusage = alloc_percpu(u64);
8814 if (!ca->cpuusage) {
8816 return ERR_PTR(-ENOMEM);
8822 /* destroy an existing cpu accounting group */
8824 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8826 struct cpuacct *ca = cgroup_ca(cgrp);
8828 free_percpu(ca->cpuusage);
8832 /* return total cpu usage (in nanoseconds) of a group */
8833 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8835 struct cpuacct *ca = cgroup_ca(cgrp);
8836 u64 totalcpuusage = 0;
8839 for_each_possible_cpu(i) {
8840 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8843 * Take rq->lock to make 64-bit addition safe on 32-bit
8846 spin_lock_irq(&cpu_rq(i)->lock);
8847 totalcpuusage += *cpuusage;
8848 spin_unlock_irq(&cpu_rq(i)->lock);
8851 return totalcpuusage;
8854 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8857 struct cpuacct *ca = cgroup_ca(cgrp);
8866 for_each_possible_cpu(i) {
8867 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8869 spin_lock_irq(&cpu_rq(i)->lock);
8871 spin_unlock_irq(&cpu_rq(i)->lock);
8877 static struct cftype files[] = {
8880 .read_u64 = cpuusage_read,
8881 .write_u64 = cpuusage_write,
8885 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8887 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8891 * charge this task's execution time to its accounting group.
8893 * called with rq->lock held.
8895 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8899 if (!cpuacct_subsys.active)
8904 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8906 *cpuusage += cputime;
8910 struct cgroup_subsys cpuacct_subsys = {
8912 .create = cpuacct_create,
8913 .destroy = cpuacct_destroy,
8914 .populate = cpuacct_populate,
8915 .subsys_id = cpuacct_subsys_id,
8917 #endif /* CONFIG_CGROUP_CPUACCT */